Trascrizione completa di #8: Build an AI Assistant Before You Build an AI Agent

over again, and probably the biggest mistake, is not giving it enough information. So a business might work in might say something like, I own a restaurant, come up with a marketing plan, or help me with my finances. It doesn't know enough about you to give a good answer. So, what AI is really, and I know we've spoken about this in other episodes, it's a prediction machine. It doesn't give the best answer, it doesn't give the funniest, it gives the one that's most likely to be correct. So if you give it more context, more information, it's more likely to be correct. So, for example, rather than saying I run a restaurant, you could say something like, I own a small restaurant with 10 employees in Meath that serves pizza. Up with a marketing strategy. It's much more likely to be correct because you've given it more information, it can narrow down the prediction and make it more accurate. The way to do that is to get it to ask you questions. So, how Niall Devitt: Okay. Marie: I always start sessions, whether it be training or mentoring, and no matter which area of AI, I would get it to say, I work as X. Please ask me 10 questions for you. Meaning Chat GPT to learn about my business. It will ask 10 questions that will be the most basic questions that you'd be easily able to answer. And that gives it enough information for it to be able to give good answers, to give good prompts, to be a reliable assistant for the rest of the session and for as long as you use that AI. Niall Devitt: Same question to you, Denis. What are the common mistakes or the common pitfalls that that people or businesses or business owners fall into when they use AI? Denis Jastrzebski: Yeah, I think it's whether it's an agent or or a chat, traditional chat bot that we've been using, it doesn't really matter. People start to use it for things that they don't understand. so it's usually business owners, you know, they're under pressure for time. They have a limited amount of resources, limited amount of staff, you know, maybe that's just themselves in the business. And it's easy to see all the social media content about how amazing AI is and it can do all these different things for you. And it can give you amazing strategies that are going to be 100% correct for you and your business. But the reality is that unless you can look at the information it's given you and properly verify whether it's a good or or a bad idea, that's the biggest pitfall because it'll tell you things with so much confidence because that's it's it's designed to give you what it thinks is the best result, like like Marie said. But that doesn't mean it's actually gonna be fully applicable. You might have missed out on a key detail. And you didn't give it that key detail, it doesn't know that it you didn't give it the key detail. So if you can't verify the result, that's a very, very big issue because it looks correct. But if I don't know, if I'm an accountant and and I'm looking at a set of financial accounts,

I can see if there's any major issues with it. But if I'm somebody that doesn't understand the double entry system, I could say, well, it looks correct because the two figures at the end are the same result, but all the stuff in between, I have no idea how to verify it. So I think that's the biggest issue is is kind of jumping into using AI for something that you have no clue about. that can cause a lot of issues and a lot of problems down the line as well. Niall Devitt: Okay. So so so what I'm hearing there, and again to kind of summarize it is the quality of the request to the large extent dictates the output or the quality of the output. But you have to be in a position or you have to know enough to begin with to know that the output, the answer that it's given you, is quality. Would that be kind of a fair Summation, so to speak. Denis Jastrzebski: Yeah, definitely. Marie: Yeah, I'd agree with that. I think some people think that AI is gonna replace things. It definitely doesn't replace thinking. It's an assistant. It can be your like like your right arm and that's really impressive. For me personally, I would only ever use AI for things I'm already knowledgeable about. So for example, if I wanted it to sketch an SEO strategy, I might use it for that because I can look at it and say that's correct. That isn't. But the example I always give is if I do a paper on, say, American history, I know nothing about it. Like I wouldn't ask it to write a paper for me on Abraham Lincoln, because it could tell me he's a zombie and I would have no way to verify it. So I think AI is best used for something you're already knowledgeable about. It's really good at creating first drafts. It's really good at create articulating ideas. It's really good at coming up with things you might miss, but it's not a replacement. It should make things better, not replace things. Niall Devitt: Okay. Okay. So AI can do potentially do lots for you. And we've talked about AI being an assistant here and helping you, I suppose, with your day-to-day work and making you more productive in terms of what you do and how you do it, etc. etc. But you know, there's that many things that it can potentially help you with. How do you decide where to start? So, in other words, You know, here are twenty-five tasks that AI can potentially help me with. How do I prioritize those tasks if I'm a business or a business owner or you know, just in terms of my kind of day-to-day role? So how how would you deal with that problem or how would you prioritize or how would you advise someone, Denis, if you were talking to them? Denis Jastrzebski: Yeah, I I always like to to try to quantify it as as much as possible. So if I'm looking at AI or even if it's just a software package for the business that that somebody is considering, I want to first of all identify what's what's the highest priority. So like you're saying, there's twenty five things on this list. One of them that I can use AI for might save me

half an hour a week, another thing might save me ten hours a week. but that that's how I usually like to look at it. So from a solo business owner point of view, that's a lot easier because you have all that all the processes for your business in your head. When you start to do it for larger teams, that's a little bit more involved because you have to coordinate with people. But as a general idea, if you can put a time a time value on it, that that's that's the best way to do it. Cause then you can even from a business perspective say, well, I pay myself whatever amount of euro per hour, maybe fifty euro an hour, 100 euro an hour, whatever it might be. And if I was able to save 10 hours a week As a business owner, that's potentially a lot of extra time that I can spend on other stuff. And it's potentially cost savings as well if you quantify it that way. So looking at how much time it can save you is how I would start, at least the priority of the list. Then you can figure out what's the best way to approach it, what the cost might be to actually use AI in that particular scenario. Maybe AI isn't advanced enough to actually do that whole task correctly. Maybe you still have to be involved in a lot of components of it. But but that's how I'd quantify it. Usually how much time can you save? Niall Devitt: Okay, so it it it's most likely then to be something that you know you're repeating, because if you're repeating it, it's taking up more time, isn't it? And it's taking up proportionately, you know, a decent chunk or a decent amount of your time per week. Okay. Marie: Yeah, Niall Devitt: So Marie: if you're Niall Devitt: again, Marie, just because you're on the the mentoring side of things. And Marie: Yeah. Niall Devitt: again, we mentioned this in the previous podcast. AI is this shiny thing, it can do so much. It feels like a bit of a kind of a black box, right? The natural instinct I would say is I'm going to get it to do something that I can't do or that I don't know anything about, right? And that's very natural. someone's on a mentoring clinic, an AI mentoring clinic with you, you know, they want AI to do something fairly sophisticated, but you know that's probably not a sensible place for them to start. How do you kind of take them back to kind of brass tacs them and get them to focus in on something that's maybe a little bit more simple, right? But a little bit more repeatable and as I say taking up, you know, quite a bit of their time potentially on a given week. Marie: Say most people don't know where to start with AI and how to get it to streamline their business. Most people don't know the full functionality of AI. Like most people don't know what connectors are, most people don't know what custom GPTs are. They usually don't have a sophisticated idea in mind because they don't know how to go about it. Typically, Niall Devitt: Okay. Marie: what they'd have is they might say they might have a department. So I would say come to the mentoring session with which department is it that you want to streamline. And they might say something like, I spent too much time in admin. I spent too much time on sales. I'm managing all my emails through email. I think I need something better. So they don't know what they don't know and they wouldn't need mentoring if they already knew because they'd already solved their problem. So what I do is I actually get AI to help them in that stage too. So once they've primed it with my business is X, you can put in a simple prompt like, I want ChatGPT to help me. Tell me what you can do for me. And it will give you the 25 ideas or the 10 ideas where it can actually add most value. And it will often come up with things that you would never have thought of. Like it can do things like help you with client communication, it can create proposals, it can create invoices, it can create a virtual reception. That's invaluable for everyone, but for particularly for smaller businesses. And they may not even know that. So they usually wouldn't even think to ask. So that's why getting the AI to advise you in the first place, I find is a really good place to start. Niall Devitt: I I think that's a brilliant example, Marie. So I think what Marie is saying there is that you you give it information about your business, right? And rather than you decide, you you you let it help you to decide and have that conversation. So I definitely think that's a very sensible place to to start. In terms of what you've seen, Denis, kind of day to day and so on and so forth, can you give me kind of some of the activities that I suppose are are are easier places for the average business or the average business owner to start when it comes to using AI and to ultimately get value, you know, get a time saving there? Wha what what battles should they potentially pick in the first instance? Denis Jastrzebski: Yeah, so because we're kind of focusing on the AI assistance part at the moment, so I'm I'm not gonna talk about agents and all the advanced stuff that they can do because it's kind of a transition period. You have to have things that an AI assistant is doing for you first before you you generally want to move into using agents. So the most classic example is it it it kind of took off in marketing first when AI first popped up on the radar, because it's one of those things that it's it's very good at at producing content. so when you're when you're creating social media posts, when you're looking at coming up with a blog post, if you feed it enough information and give it enough instructions on how you'd like that writing to sound, how you want to come across, that can save a lot of time because a lot of people aren't confident in writing. writing is is a big, big skill and most business owners have a million things to do. And to sit down and try to write a blog post or to write even five social media captions if if you have a million other things to do, it's it's nearly impossible

because it's hard so hard to be creative when you're under that that sort of pressure. So marketing, there's so many functions within marketing that AI can assist you with. So I think that's where most people end up starting. and that's where a lot of businesses tend to fall short is they're so busy that they don't have time for marketing. And then when they get quieter, then they spend some time on marketing, then they get too busy again and stop doing marketing. And it's a constant cycle whereas with AI, If you start using it even when you're busy, you could potentially have a constant, you know, marketing and sales cycle, which would be ideal. So marketing would be the first one. And then admin being the second one tends to be a very popular thing. It's repetitive. It's something you have to do often. maybe you have to draft a bunch of documents, maybe you have to to go through lots of email replies and and take out the information that's really relevant. there's a lot you can do with with AI with you know exporting emails that you have to to go through. If you had to read through them manually, you can upload them into AI, help you organize your thoughts, help you do brainstorming, figure out what has to be done from the admin side. You might have some sort of little bookkeeping question and you can't really figure out or you can't remember from the course that you've done, how exactly do you treat this particular little transaction? And it is really good for that sort of stuff because it's pretty predictable, it's pretty structured, and you have to do it often. Marketing and admin being the two most common ones there. Niall Devitt: Okay. Okay. So okay, we we we've talked about there, I suppose, you know, if you're unsure where to start, how you can actually use the AI to help you figure that out. Right. We've talked about maybe some of the common tasks that you know a business can potentially look at, right? But let's just say we we've identified where we want the assistant to be, where where we want the assistant to help us with. Right. So how do we then go about creating an assistance that will offer value, that will offer productivity, that will offer time saving to us? Yeah. So is there any kind of structure or rules? Is there anything that we need to follow there that will get us there on purpose rather than kind of by accident. So in other words, Marie, what ultimately at the end of the day makes a good AI assistant in your opinion? Marie: To get AI working well for you in general, it starts with the prompt. Prompt is, and it's just a fancy word, it's where you tell it what to do. You're basically poking it with a stick in a certain direction. There's lots of different ways to get a good prompt, like lots, and we'll talk about them shortly, but I've found that the easiest way is to get it I I'm just a big fan of getting AI to do as much as it can for you. Get it to write the prompt for you. So Niall Devitt: Okay. Marie: the term that I

use, and it's a game changer for a lot of people, it's act as a prompt engineer. So let's go back to the example of SEO that I mentioned earlier. If I was to ask it a prompt for SEO, I know about SEO. So I will mention things like keyword research, backlinks, pr, digital pr, because I know to include those things. But if I don't Niall Devitt: Mm. Marie: know enough about that, I wouldn't even know where to begin. So instead, I would say something like, I want to do SEO, act as a prompt engineer, and create a prompt. So it creates the prompt for you. You can use that in any area of your business. You can say act as a prompt engineer and create a prompt to help me with finance. Act as a prompt engineer and create a prompt to help me with admin. Niall Devitt: Okay. Marie: Once you get used to that, you can say something like create 10 prompts for admin. It will create 10 prompts, and you can, you the human, the user, you can decide I like this one, that one's useful, that one's a complete waste of time. And then you can get it to run the prompts for you. so yeah, as I say, act as a prompt engineer. That's absolutely powerful, particularly when you're dealing with something new, or when you're new to AI or you're using a new AI tool. Like if you've used ChatGPT for a while and then you want to branch out to Claude or Copilot, I would still stick to Act as prompt engineer for a little while. Niall Devitt: Okay. Yeah. And and I and I I think even when you know how to prompt and you know how to prompt well, I would still I suppose recommend that that's still a sensible thing to do. So in other words, no one probably understands AI as well as AI does. So even when I've created a prompt, I will always ask the AI to review that prompt and help me to improve that prompt before I use that prompt, depending on on on kind of what it is for. Right. So Marie has mentioned the prompt there. She's mentioned how we can actually use the AI as a prompt engineer to create a better prompt or help us with our prompting. What other things do we need to be aware of, Denis, when working with AI? What other steps do we need to take? Denis Jastrzebski: Yeah, so it's it's very easy to get started with all those things because then suddenly you have a really, really strong prompt and you feel kind of a little bit more confident about the output. but two things that the end up being either an afterthought or skipped altogether is is setting up different kind rules around what it should and shouldn't suggest and kind of how far it should go with the information that we're we're looking to kind of assess or maybe the type of strategy it's trying to come up with. So having rules around that. So you can include that in the prompt. You know, you can, as part of the prompt engineering, you can get it to create some rules, at least the a beginning set of rules that'll help it stay within its lane, not to kind of go beyond stuff that it might start to hallucinate or come up with things

that aren't maybe as relevant. So kind of give it a boundary. Niall Devitt: Okay. Denis Jastrzebski: once you do that, then you can be a little bit more confident as well that if you're giving it a prompt about coming up with a strategy, it's going to focus on the strategy. It's not going to go beyond it. so rules are really, really important. But then at the Niall Devitt: Right. Denis Jastrzebski: very end of the whole process or even during that process, reviewing what it's doing and what it's suggesting and then figuring out why it's suggesting those things have to be at at the the kind of core of all of this because we shouldn't let AI come up with a strategy, come up with all the let's say a marketing strategy come up with all the social media posts, and then without a second thought, we just copy every single post. paste it into our our social media scheduling tool and then just fire that off without reading through any of it. We need to have human review as part of any anything that AI is doing because it's not going to catch things that you know about your business that you haven't given it yet. It's going to make up information sometimes. So you even with a really good prompt, it has a tendency to to come up with things that don't exist or suggest things that don't exist. It's a common problem with every single AI out there at the moment. and as a general rule, if we don't review things, a lot of a lot of things can go wrong. So having human review as part of any process, whether it's marketing or admin, you need to keep those two things really central to whatever you're doing. Niall Devitt: Okay. So so Marie has mentioned prompt and prompting and how to do that. we've already mentioned context, we've also mentioned rules or what sometimes are are called constraints, and we've mentioned review, which certainly when it comes to AI assistants and as you're getting used to AI, to my mind has to be led by the human being. Here's what my experience has been. When I've got AI to do stuff that's genuinely valuable for me. And by genuinely valuable, I mean, you know, has saved me quite a lot of time in relation to something that I've done. What I found to date is it's also a process of iteration. So, in other words, with all of that stuff, right, I rarely get it exactly right the first go around. Right. So it produces something that's probably 80% of the way there. Right. And I need to figure out to get how do I get it from eighty percent to ninety percent. And oftentimes that's kind of stuff that that maybe I hadn't thought about at the start. So I suppose to give you a very practical example of this, right? So Denis would be familiar with this. So I I was creating a custom GPT, which is basically a kind of a trained version of GPT for a client. there recently in terms of the Marie: Yeah. Niall Devitt: planning and to a certain extent the creation of their content. So first time around it create it it it created pretty good content, right? But

it wasn't referencing certain web pages on the website. And what it was also doing is it was repeating content that was already available on the website. So so I suppose kind of two rules or constraints that I had to build in there was when it was creating content or creating a content plan that it would point to wherever that expertise, that service, that product was on the website. And if the information already existed on the website, there was really no need to go into it in detail again. But as I say, when I was creating the assistant in the first instance, those weren't things that I was necessarily aware of. Now, ironically, when it came to the client, it created so we we got it to create a s where it was creating, you know, really good content with their review, but then they asked me for my opinion. And the issue at that stage is I didn't know enough about their industry, I would say. And I had to go back to that, look at fine to me, but I don't know enough about your industry. your customers, your products, all of that type of stuff, you need to review that content from that perspective. So something that you mentioned before the podcast, Marie, but I do think it's very, very important. And again, interestingly enough, you didn't just talk about role, which I'm going to ask you about now. You talked about roles plural, which is something I haven't done with AI. So it was an interesting one for me. Talk to me about role. What is role? When it comes to AI and why is it important? Marie: Sure, okay. So most people treat AI like an encyclopedia or like Google search. They ask it a question and they get an answer. But AI has so much information. If you give it a role, so give it something to do, pretend you're a director, it will act like it. It just make it fun. So it could be anything Niall Devitt: Okay. Marie: like, pretend you're Gordon Ramsay and give me a recipe for scrambled eggs. It could be anything like that. In business, you can give it a specific role. Now I found it really hard to get my head around it because a lot of people mentioned role, but I'd never really seen it explained. So one day in a session, I decided to give it three different roles. So I picked create a marketing strategy for a gym. And the three roles that I gave were intern, marketing strategist, and accountant. What I was expecting to get was obviously to get three different answers. But to be honest, I was expecting to get an absolutely terrible response from the intern, a terrible response for the accountant, and of course a brilliant one from the marketing strategist. And then that would make it really, really obvious why role is so important. Instead, what I got was three completely different outlooks. So the intern was very focused on quick, fast now. It came up with a really good strategy focused on TikTok, and it was talking about a younger audience, which is correct. Marketing strategists would be my favorite. They were talking about long term, talking

about persona, talking about email marketing and Instagram, which was also correct. The accountant's one was fascinating. He was talking about it was talking about how basically not to spend money, but how to get more business out of your existing customers, how to increase their lifetime value, and how to use them for referrals. All three were really, really good. And unless I was doing that exercise, I would never have branched out from act as a marketing strategist. So as well as just giving it a role, giving it a correct role, you can get it to give you the perspective from different roles altogether. And if you work in a smaller business or if you're part of a smaller team, we all have our own very narrow lens. Like even the most open-minded person has a restricted viewpoint. AI does not. It has basically everything that's been written throughout human history. It's incredible. So draw on it. And in that way you can get different perspectives from different areas. And again, I'm always a really fan, a big fan of getting it to do things for you. So you might say, I'm considering this thing. Can you suggest three different roles to look at this thing that I want to solve? And it will give you three roles. Then it will give you the perspective from each one. And from there you can think, well, that one's really interesting. Maybe that isn't. And it'll help you come up with a more thorough Plan. It's like having a small board of experts on your team all the time. It's really powerful. Niall Devitt: Okay. And and a role that I think is is probably useful for most, I suppose, asks with AI, Denis, and you can agree or disagree with me, is the role of the customer or customers. Right? so I think Marie mentioned there that you can give it the role of an expert or a professional or you can give it multiple roles, multiple experts and and multiple professionals. But to me a good role to have in in in most circumstances is the role of the customer. Right. When it comes to roles in terms of what you've seen, Denis, is there anything there that we're overlooking, we're missing out on? Is there you know, is there anything to I suppose more generally, how how do we make that whole process of getting to an assistant quicker, you know, faster, better, more effective, more productive? Denis Jastrzebski: No, I think what what you both had covered there it really is on point. I wouldn't add too much to it except you can kind of go I suppose with with an alternative way of doing it. n one way versus the other isn't necessarily that different Niall Devitt: Yeah. Denis Jastrzebski: in terms of the result, but here's just an alternative way. So when we're looking at our businesses, whatever we're getting AI to do, we can stick with the marketing example. When we ask it to come up with a strategy, we have our prompts all set up correctly, we have all these other rules and and and so on. So, you know, we're coming out from our own perspective and we're saying, well, come up with a strategy based on these details that I've given you. but a lot of businesses, as much as they don't like to admit it, it's rare where there's a business, you're a business owner or you're running a company and you're completely unique in what you offer and you do. It's practically nearly impossible. and if it is at the moment, there's probably going to be a competitor that'll pop up in the next 12 months, right? But most people don't want to to necessarily admit that that their product or service isn't as unique as it actually should be in their mind. But I like to always look at competitors for a particular business when when you're doing anything with AI. So whether it's like automation, whether it's marketing or admin, I like to get AI to research. what our competitors are doing in in the particular area that I'm looking to improve on and use that as the context. So I'd say, you know, visit their website, try to identify what their USPs are, so what their selling points are. What the what does our competitor focus on as their kind of main selling point? What's their kind of marketing strategy like? What do they tend to repeat in their messaging on the website? I might feed it five or ten captions from some of their social media posts. So once I give it all that that information, I can get it to analyze it and pick up pick up on stuff that I wouldn't have picked up on. and that's kind of the whole idea I've I've used from from the beginning is market research is one of the most underrated things that businesses tend to not have time to do on a continual basis. And a that's Niall Devitt: Hundred percent yeah. Denis Jastrzebski: where AI can really, really help you. And it it's it's stuff that you can set up and it'll continually pull in new information from your competitors, which is a really, really big help. so market research, I think you can start there. So look at your competitors, pull in information from your competitors into the chat, and then see, you know, what are they doing that I'm not doing? So instead of trying to reinvent the wheel and coming up with something brand new, it's better to just see what why are these competitors doing better than me and then what can we kind of improve, or maybe we can come up with a version of that for ourselves. maybe they're selling subscriptions. Well, we're not, and it seems like they're a lot more successful. so that way you actually have a lot of the thinking done for you and then you can just implement stuff. You can actually look at how your competitors are doing that thing. so it's just an alternative way of starting that process, but you continue with the same prompts with the same rules, you just start with a different set of information at the start. Niall Devitt: Okay, fantastic. So a couple of caveats, and I've had the experience to a certain degree. So Marie is absolutely right in terms of I suppose the available information out there. However, I have encountered areas

when using AI when the information out there isn't so good. So I've had to give it new and better information, right? So myself and Marie from Tipperary. we're probably both hurling. Supporters, but I did do a project a little while back where I was trying to train or build an AI to help me with training session plans for my young lads hurling and football team. When it came to hurling, its understanding was very limited and very wrong in certain instances. That's because the information that's out there in terms of hurling is a little bit confusing. Right. So what I had to do was I had to give it some knowledge in terms of hurling and how that works, and specifically in terms of trying to train or coach young lads. So it would j it just wasn't sufficient. That was quite a challenge for me to do, right? The other thing I would say, and again, this this point is true in terms of businesses, right? You know. Oftentimes the information that's on our website, etc., etc., is so so in terms of the business, right? The information is either sitting on your hard drive somewhere, or in a lot of cases, it's in here, right? So I would say when it comes to your business, and we talked about context earlier on, right? And certainly something that I've experienced in terms of AI, the better the context, the better the information that you give it about your business, the better. the assistant or the output will be. And oftentimes that information is not available to it, or the information that is available to it is, I suppose, you know, weak at best. Talk to me. And again, broadly speaking, Denis, as I understand that what this is, what this is known in terms of AI is is knowledge base. So we're into kind of knowledge base here, but but talk to me about that, please. Denis Jastrzebski: Yeah, so whether it's an assistant or an agent system, both both of them have to have some sort of information that we we've given it as as that kind of knowledge base as you're saying. So we talk about the knowledge base a little bit more in the context of having AI agents and starting to move into that side of things. So I'll touch on that as well. Your knowledge base really is everything that you want the AI to know. But not only that, it's in a very specific format that the AI can understand. So you can't necessarily just upload a bunch of different PDFs and random files into your knowledge base. You technically you can, but that's not the most efficient way and the best way for AI to actually can interpret that information. So you you tend to have it in certain formats, usually in what we call markdown. But your knowledge base is the key to anything that you're doing with AI agents or with assistants, because with an AI agent If you imagine it now that we're asking it to do things on our behalf, there's a lot more risk, a lot more things can go wrong. So, you know, if I'm just drafting an email in my chat GPT window and using my assistant for that, yes, it can say something

that's incorrect, but I can catch that and change it before I send the email. But if I'm using an agent and I'm saying, Well, draft that email and then send it, and I don't want and I don't put myself in the middle of that process to review it, then the risk becomes a lot bigger. So if there was something incorrect or it came up with some sort of random piece of information in that email that it sent, we didn't catch that. So your knowledge base is is both everything that you want AI to know about your business, but it also includes all the constraints or as we call them guardrails in what you want it to do. So if it's sending an email, for example, to an existing customer, you can have a rule that it'll never send that independently. It'll always run it by you because you don't want to obviously ruin a relationship accidentally. Whereas, you know, if you get a very common FAQ query from your email, somebody's asking the same question that's popped up on your website that you have it as an FAQ, you can pretty confidently say just use the information exactly on the website to reply to that customer. Now, I would generally say you still want to review everything before it goes out to customers or any kind of public-facing information. But in theory, you could do that. and there's a lot less risk. So your knowledge base includes all the information. But also all the rules that you want it to abide by when it's doing things on your behalf. So it's a really kind of key component of of any sort of AI system that you have to kind of be a aware of, yeah. Niall Devitt: Okay, okay. So I suppose a kind of a simple framework then that we could probably recommend when it comes to using AI is to give it a role, or as Marie has pointed out, potentially roles, to give it a task or an ask, right? To give it context or information, to give it rules or constraints, what we want it to do, what we don't want it to do. And for there absolutely to be human review built in, and whoever that human is, they know enough about the subject or the subject matter or the task to know whether it's done a good job or a bad job. I would probably include, particularly if you're trying to get it to do something that's you know semi-meaty or semi-sophisticated for me. you know, it's it's it's not a one-step process. It requires iteration. So I suppose the example I would kind of give here is is we're using a custom GPT. So again, just before we kind of move on with that, do you want to just give a a brief Marie: Yeah. Niall Devitt: definition of what that is for someone who's not familiar, Denis? Denis Jastrzebski: The custom GBT, yeah. So that's that's a very, very basically just a customized version of a chat GPT conversation that you're having. So if I open up a new chat GPT window, it kind of starts fresh, but there's also stuff from my account that it might know about me. But that's your kind of regular window. A custom GBT is you you've given it a an additional set of instructions and maybe even some files that have popped in there initially. and then it'll always answer you based on all those files and instructions. So that's all all a custom GPT is. but that is changing. So OpenAI, the company that has released Chat GPT, they're actually getting rid of that feature. So they're changing custom GPTs. most customers at this stage in September of 2026, that's starting to disappear from most accounts, but it's getting turned into something slightly better. in a way, but OpenAI hasn't really maybe publicized that too well. So it's it's becoming what they call plugins, which is the exact same thing. You've got a chat window with additional instructions, some additional files or a kind of knowledge base that we've we've talked about, but it also lets you connect some tools to it. So you can actually connect Gmail to it as a source of information. So it can read all my Gmail emails if if I wanted to, it can view them. And it can bring that context into my custom GPT, which can be very, very useful. If you start thinking of all the different assistants that you could potentially build, maybe from a customer service point of view, that's quite handy because instead of having to copy and paste an email that a customer has sent you and you're kind of trying to draft something that's nicely, nicely set up, you can actually get it to just pull that email straight into the chat for you. so it's becoming chat GPT plugins. But it's it's essentially the same thing, just a little bit better. Niall Devitt: Okay, so the the example kind of simple example I would use here is so when we were starting off this podcast, I created a very simple custom GPT for the podcast, just in terms of the episode planning, right? So t it was to help to to put episode plans together, to to share them with you guys, to get your feedback, to improve them, etc. etc. etc. Right. What you like, what you don't like, topics, etc. etc. Right. But what I've been able to do since, depending on your preferences and stuff like that, is I've been able to further customize it. Right. Not only that, I've been able to give it, I suppose, the how the podcast is performing. So what is the audience responding to? So again, as I say, there's a process of iteration there, and each iteration makes that assistant better so it can do. Ever so slightly more, right? Which is more useful in terms of kind of what we do. You spoke the last day in terms of how we're using AI in terms of the the the editing and the production side of things and then the promotion side and so on. So but again, you know, the key point here is there has to be the human review. I mean steering the ship in the first instance, but then also putting that out to you guys and getting your feedback. you know, so it's not a kind of a one-off. Okay, I want to move on and I'm gonna go back to mistakes again, right? But I want to delve a little bit deeper this time around. And I suppose when it comes to

creating an AI assistant, and I or if we can in as much as we can, right, I want to look at mistakes in approximate order, right? So there's kind of mistakes that we make along the way, or we potentially fall in. And I'm sure we've made these mistakes as we were kind of learning how to do this. But I suppose to give the audience or the viewer a steer in terms of kind of doing that. So again, Marie, going back to kind of the top mistakes that people make when it comes to AI, and again, you know, building an assistant, something that's going to kind of add productivity, add time saving, etc. What would you say is the number one Mistake in terms of order. Marie: in terms of where they start. Niall Devitt: Yeah, yeah. Just in terms of the kind of order, the approximate order in terms of mistakes. I I'm gonna give you I I I'll tell you mine in a second, but I'll let you go first. Marie: In terms of order of mistakes, I think yeah, they get a bit too much information at the start. And we've all done that where we go down the rabbit hole and AI can do this and it can do that, and you're like having this long conversation. And ChatGPT is kind of notorious for it where you ask it a question and it gives you the answer and then there's a follow-up question. So it tends to go into a big long conversation. So at the end of like an hour, you have a ton of information, but nothing to do with it. one way around that that I found is called the Pareto principle. So I use a prompt called 80/20, where so with the Pareto principle, you get 80% of your results from 20% of your effort, but you need to do 100% of the the task to figure out what the 20% is. So at the end if I have a ream of information, I'll ask it like to do 80/20 and tell me what it is that I need to do. Niall Devitt: For me, the first mistake, Denis, and this is one we dealt with in episode number two, if I'm not mistaken, is people starting with the the technology, the AI, rather than properly defining the problem or properly defining, I suppose, the task. That's where I suppose that's the first trap I think that people fall into. Now again, I think Marie had a good solution there in the sense if you go to the AI and you actually get it to help you with that. But in the absence of doing that, is that something that you see? Denis Jastrzebski: Yeah, I think it it it becomes one of the yeah, I think it's it's probably the top of the list in my head, at least from the perspective of the work that I'm doing. And the Niall Devitt: Yeah. Denis Jastrzebski: the reason for that is it's not anything new that that's been the issue with AI, or it's not because AI came in and started and people just don't know how to use it properly. This has been an issue for as long as technology has existed. Companies, whether Niall Devitt: Hm. Yeah. Denis Jastrzebski: they're big or small, invest in platforms and tools and in consultancy that could cost hundreds of thousands and dozens of millions depending on the company.

And they implement something because it just has all the bells and whistles. Somebody sold it to them and it looks and sounds perfect. But they forget all the other more important stuff, which is bringing the people in the company along which you're on the journey, kind of reassuring them about how they're going to use the software, training them on how to use the software. All that stuff that's always existed. So companies invest millions, then you know, Stacy in the accounts office ends up creating a spreadsheet that does it better than than the tool that has millions of euro invested into it. So and everyone else in the the the team also preferred the spreadsheet. and that becomes a big issue, a lot of money spent and a lot of time spent on something that wasn't really needed. And with AI, it's exactly the same. It's very easy for for companies and for people to To try to s start with using AI because everyone's talking about it before realizing or kind of even evaluating what they actually should use AI for, that's that's useful. Suddenly they've you sunk a bunch of hours into creating prompts and all these different really, really detailed processes, but then they realize like they don't even follow that process themselves. And the staff that are working on it don't follow the process themselves. so that's that's the biggest issue. People people start with the technology because it's new and exciting. But at the end of the day, it all has to start with kind of defined processes, very simply doing the boring work of documenting what a current process is. Maybe there's some scope for improving it as well. But even if we just have a current process, how can we use a little bit of AI to make our life a little bit easier in that? That's where people should start, as opposed to to getting something really big and shiny and then throwing throwing it at the wall and seeing if it sticks. it just doesn't work. And it's a lot of time and effort spent for for Not much of a a return in my experience. Niall Devitt: So I think that's an important one there, in the sense that and we kind of alluded to this, how important process is, or having a process that you can explain before you attempt to get the AI to help you with that process, and and that's definitively because. The trap that you will fall into is if that process isn't well defined and you can't explain it properly, AI will invariably make the situation worse and and and waste quite a lot of your time. Another mistake, I think that, and and and definitely we've kind of covered off this earlier on the episode and and re talked about it, is is substandard or vague kind of prompts. So, you know, you you your ask has to be. specific, it has to be quality there. And again, if you're not quite sure how to ask, you can certainly get the AI to help you with that. I think that's a really good way to do that. I mentioned treating the first answer as the finished answer, so to my mind, you know, you've got to accept, particularly if you're looking for something that's sophisticated or valuable, that there will be a certain amount of iteration involved, so you'll improve as you go. As you see the outputs, as you test the outputs, et cetera, et cetera. Something that we've kind of alluded to before, right? Definitively for me, and I've fallen into this trap myself a couple of times, and it's where I've gotten unstuck with AI is asking AI to do work that you cannot judge or cannot do yourself. I think that's a a definite one. one for you then, Denis, because again this is something that you looked at earlier on in one of the earlier episodes. When we're giving the AI information with respect to our business, do we need to be careful there in terms of the information that we give? And what information should we give? And perhaps is there information that we shouldn't give? Denis Jastrzebski: Yeah, there definitely is. And especially if you're using a free version of ChatGPT or any other model, that becomes infinitely more important. Even if you're on a paid version, there's still a lot of things to be aware of. So unless you're on an enterprise version of AI, which you're not unless you're in a larger company, you have to be really careful what what you upload there because by default, in every single AI platform that you use. There is a small notch in the settings that says anything that you upload to this platform, we can use as part of our training to make the AI chatbots better. And you think, well, that's kind of harmless, you know, nobody's really going to read those. But if you're uploading any sensitive information about your company, or even more importantly, about your clients, if you're working in that sort of a business, that can have a lot of repercussions. So, but from a GDPR point of view, there's a lot of things that are being violated there. from a client confidentiality point of view, a lot of issues. and you might think, well, it's kind of going into this this system, it's going to be used maybe for training, and then it doesn't really go anywhere, it gets deleted, it says, you know, this will be deleted after a certain amount of time and no harm done. But if if OpenAI or any of these companies were ever hacked, there was a leak of any sort of information on the servers, there's potentially your information and client information floating around on the internet suddenly. being sold on the the black market even. So that's the worst case scenario. That's if you don't get caught and at the very, very end that's what happens, you know, then it's a disaster. But everything in between is also going to have repercussions. So we never really want to upload any sensitive information into our chat window. So if you're thinking, you know, what what would be class of sensitive information in the context of what I would use Chat GPT for if I'm trying to understand my financial accounts. if I'm uploading my my account from my accounting package or

what my accountant sent me, just though it is straight into ChatGPT and I'm kind of asking it, well, could you explain what what all this stuff means? Where should I be looking at? You know, maybe what are some cost savings? You know, it would you be happy with taking that and sharing that on your social media? if you weren't, then it's probably not something you should upload into Chat GPT because if it ever does get leaked, if it ever does come come out into the open, there's lots of issues around. and then the the other part of that generally is that if you're uploading any of this sensitive information, even though it might not get hacked, it is getting assimilated into this massive training data of of all these big companies. So it might not directly come up in you know conversations. It's not like it's gonna pop up for other people in conversations, but there is a chance that especially if you're working on something really unique, if there's some proprietary information. that can get assimilated into you could say the big AI brain for that company. And it could potentially be used for you know indirect purposes in other conversations that users have with it. So there's all sorts of issues that pop up. but it is something to be very careful of. And you can just very simply anonymize anything that you're going to upload that you think is sensitive. So you know, if I'm uploading a set of accounts, I'm gonna take out my company information, anything that could identify that data with me or my client. Because now if I upload that, that's just a random set of numbers on a spreadsheet, some sort of accounts, but nobody can link that back to me. So that's a an easy thing you can do as an example. But generally, you know, can I have a have a think whether this is something you actually do need to use AI for? Maybe it shouldn't even be something you use AI for in the first place. Niall Devitt: Okay, I think we've given it a pretty decent overview there, and we've started to put some meat or some detail on the bones in terms of creating an AI assistant that will be productive and time-saving in terms of the business. So, like we've done on the other episodes, just before I finish up, I want us all to recommend one tool, one technology, one approach, whatever it it's it's going to be. So I'm going to start with yourself, Marie. What would you recommend to businesses or business owners when it comes to creating an AI assistant? Marie: Number one, get it to ask you questions to learn about your business. So that's what I would call priming. Number two, act as prompt engineer. It is powerful and it's a term that they all understand. And then number three, the Pareto principle, 80/20, when you finish, just to give you the actual the three or three to five steps that you need to do whatever it is you need to do. Niall Devitt: Okay, so three there, and that's something that we'll kind of publish with the with the with the show notes there. Just a quick question, Denis, we've talked about I suppose, you know, knowing enough about whatever you're asking it to do to know that it's done a good job or not. But how do you measure success when it comes to AI? So you you you know, the output looks good, I know it's good because I know enough about it, but how do you actually know? That it's been valuable in terms of my business. What do you measure? Denis Jastrzebski: Yeah, well it it entirely depends on on what you're asking it to do, but I think I'd like to link it back to step one, which is when we're identifying what we're going to use AI for. And we're saying, well, this can potentially save us ten hours of our time. That's what we're assessing. once we actually start to build out that assistant, maybe we've gone through iterations like you said, and now we're starting to test it on on the actual day-to-day Niall Devitt: Yeah. Denis Jastrzebski: task that we've we've kind of designed it for. we should Keep track then of the time that it's actually saving us. So if we we look at our clock and we say, okay, this usually took me half an hour. I'm gonna set a timer and I'm gonna use it with the assistant that I've built. Then you stop it afterwards and you say, okay, I actually only took 10 minutes this time. And that's included with maybe little adjustments or revisions that I've given it as well. So then you can start to quantify that from a time perspective. Then you know, okay, it is saving us the kind of time we expected, or maybe it's only saving us six hours instead of the 10 that we we had estimated. But but you should then really kind of spend a bit of time on figuring that out because if you just kind of build the assistant and then let it go doing its thing, you work with it, you might not realize you could potentially increase the amount of time it takes to do tasks. and you might not necessarily realize that either. So it's important to track the whole kind of loop as it were, with with your time. And in other yeah. Niall Devitt: Yeah, I I I I I I would say it's absolutely critical because you don't know, right? Denis Jastrzebski: Mm. Niall Devitt: And and again, so again, if we take the podcast one as the example, we were able to estimate that without the kind of AI involvement, you know, there was probably a collective about fifteen hours per episode between myself and yourself. It's five hours of our time, so it's saving us ten hours. But y you need to measure that. You need to track that. So I think time in the first instance is a very straightforward one to measure. But I think it's absolutely critical to measure it because you don't know if it's actually adding value or productivity or not. In terms of your recommendation then, Denis, what are you recommending as an approach technology? Denis Jastrzebski: So we've been contrasting agents and assistants throughout the episode, and we've obviously now figured out a way and a nice framework of how to develop assistants. Well, Marie has

some really good suggestions, and then you've brought that in Niall. We kind of have a nice maybe six or seven step process now. But once you're moving into, you know, thinking of what should I build an agent? once you've defined that process, you've built an AI assistant and you found okay, it's actually saving us that amount of hours, it's working really well. Well, you could probably optimize it further. That's when you should start considering maybe an agent could also be included in this process. Maybe it can take over this process and do even more stuff. But it's a very complicated process from there. So I'm sure we're gonna have a dedicated episode on agents. And of course, Niall Devitt: Yeah. Denis Jastrzebski: that example that we're gonna talk about in a future episode, I'll be interviewing a business that we we've worked with. but I wanna set everyone up with kind of a really nice baseline. So if you're starting to think or even want to just mess around with AI agents. without actually using it for for the business side of things, that relies on the knowledge base that we had mentioned earlier. So there's a particular file that regardless of what kind of AI agent platform you might use, that could be like Cloud Cowork, it could be Gemini Spark. They all work on the same kind of principles. There's that base and set of instructions, that knowledge base. And within that, I've I have a file that I use in my own system that helps keep everything in check. It documents new preferences and new rules that you you put in while you're working through all these different processes. So I'm going to mention that and I'm going to have a little set of instructions in the show notes on how to actually set that up in whatever environment you're using. And from there, that'll be a solid base to just start testing and experimenting with AI agents. And if you have a successful assistant, you might want to then try to build that in in an agent environment instead. So it I think that'll be really useful and I'll include that in the the show notes for everyone. Niall Devitt: Yeah, so I'm very going to quickly recommend an AI that I'm sure some people have heard about and others possibly not. Somewhat controversial because it was it was initially bought by Meta by Facebook. so it's an AI called Manus AI, and I'm going to recommend it specifically for one thing, and that's building websites. Now I should probably Marie: Yeah. Niall Devitt: qualify this by saying it's a particular website. type of website. So as I mentioned on the previous pod, I'm involved in another kind of a hobby podcast with my mates. We wanted to create a a website for that podcast. I probably know enough about WordPress that I could go on there and create a website. It would still take me some time if I was to do it to kind of any reasonably decent standard. Manus AI has a nice feature where it will build websites for you. Now I've only built one website using AI. As I say, it was a kind of a hobby website per se, but it did a fantastic job. It didn't do it in one go, it required a certain amount of iteration. I know a little bit about kind of building websites, but certainly if you were kind of a startup or a very new business and you wanted to get a website up. quickly it's one that I would recommend. We've gone a little bit all over on where we'd like to be. So before we finish up, and I think we, you know, we we we got into the meat of it there. So I'm very happy with that. But again, just maybe some final words or some final takeaways. So starting with yourself, Marie, any final words or takeaways? Marie: Yeah, sure. So again to reiterate, the best thing you can do with AI is get it to tell you what to do. So start by getting it to ask you questions and then ask it what it can do for you. and then once you get the feel for it, you can kind of figure out what things it can actually help you with. And then it might be you might find that using it as an assistant is enough for most people. Beyond that, you can get to grips with things like cloud connectors projects and custom GPTs and then obviously migrating to plugins. But start with doing the basics really well first. And as I say, that might be all you need. Niall Devitt: Yeah, yeah, I totally agree. So and again I say I I reiterate if it's just as simple as time measure because once you start to see it's saving me five hours per week Suddenly you get to switch, you get to start to get switched on in terms of the potential of AI, right? So as I say, it's not just a case of using it, it's also a case of measuring the output and proving to yourself that it's valuable, it's productive, etc. etc. Denis, final words or takeaways from you. Denis Jastrzebski: Yeah, final reminder is if you're transitioning from AI assistants to AI agents, it's even though it sounds and looks like a very similar approach and system, the way we interact with agents and assistants is is quite different fundamentally. So with assistants, it's everything everything is about the way we prompt and the instructions we give it and so on. With AI agents, you you pretty much reverse the process so you don't have to do prompt engineering. You don't have to do really, really complicated instructions and set up all these different rules and guardrails initially. What you want to do with agents is you want to give it the end result that you're looking for and let it figure out how to get to that. So if you try to give it a set of really detailed instructions from a prompt engineering prompt that you've designed, it's going to try to fit into that, but it might not do it very well. It's going to know much, much, much more efficiently how to do it itself with a slightly different approach. than than we ever will because of the way it's designed. So it there's kind of two very fundamentally different ways of approaching how you work with agents versus assistants. So that's one thing to keep in mind. You'll see that in the instruction file I mentioned in the session notes. but it's it's a big distinction. Even though it looks like it should work the same way, it's very different. Niall Devitt: Okay. So just to reiterate for anyone that's new to the podcast, so Denis does up a resource page with every episode. So that has all of the resources that are mentioned during the the episode itself. So that's a fantastic place to go. We publish it on the content plan website. the podcast is available on all of the main podcasting platforms. If you're enjoying the podcast or if you've enjoyed this, we'd really, really appreciate. That you subscribe and even review, folks, that would be fantastic. Can I thank you very, very much, Denis, for your time and your advice today. Marie, fantastic to have you on board. It's taken me a few episodes to get you on here. Up tipp. Thank you so much for your time, your wisdom, and your advice today. and bye from us a content plan. We'll we'll see you on the next digital and ai decoded episode, folks. So thanks very much and bye bye. Denis Jastrzebski: Thanks, Marie. Thanks, Niall. See ya. Marie: Thanks everybody. Bye.