20 | No-Code Superpowers: Unleash Your Productivity Harnessing the Power of Free Automation Tools and ChatGPT with AI expert Pierre-Louis Guhur

0:00Hello and welcome to Leveraging AI. This is Isar Maitis, your host. Today is a more advanced and highly practical and effective episode. we are going to talk about how to infuse AI capabilities like ChatGPT together with automation tools that already existed before in order to create absolute efficiency magic. Now, if you feel this may be a little too advanced for you, I suggest you start with the previous episode, episode 19, that is more of an introduction to AI, and then maybe join this one. But if you have any technical skills, and definitely if you use automation tools like Zapier before, you are going to absolutely love this episode because it is going to take your automation game into a whole new level.

0:44As always, at the end of this episode, I'm going to give some exciting news from this past week. And now to this amazing automation episode. In the next few years, AI technology will change our world dramatically. Whether you are a business executive trying to catapult your business forward, or just somebody who refuses to be left behind and want to advance your career, this is the show for you. I'm your host, Isar Matis, a serial entrepreneur and an AI enthusiast. You'll hear invaluable practical tips from innovative business leaders, AI practitioners, and some of the brightest AI minds in our world today on how you can leverage AI in ethical ways to advance your career and grow your business.

1:36Hello and welcome to Leveraging AI. This is Isar Maitis, your host, and I got an amazing episode for you today. My guest today, Pierre-Louis Gouhour, he is, first of all, French, as you can imagine from his name, but he's literally one of the smartest people I've met in the past few months, and definitely one of the most advanced when it comes to amazing implementation of AI tools across multiple workflows and use cases, from personal ones to he's teaching himself Italian in a tool that he built. And we're going to talk about this with him shortly. But in his background, he's a successful serial entrepreneur, and he has an advanced degree in AI, which makes him the perfect person to do all these kind of magical things.

2:18He's probably about to change the world when it comes to adapting houses to climate change in France, which is what he does for a living. But today, we're going to talk about how to identify processes that you can automate using AI and additional tools in your company. And we'll take a couple of use cases and actually walk you step by step on how you can do that yourself. And it's really, I've been spent over an hour with him going on all these use cases. It's mind blowing what you can do on your own today without writing code or with minimal code or scripting. And we're going to walk you through that today.

2:57So Pierre-Louis, I'm so excited to have you. Welcome to Leveraging AI. Thank you, Izar, and thanks a lot for your very kind message. Let's start with the what. One of the problems I'm sure many business people are facing is how do they pick the right processes they want to address with AI? Because AI can do all these different things, right? So you can create code, you can create automations, you can write stuff, but it touches almost every aspect of a company. What is the right process for us to figure out what are the lowest hanging fruit, what we should start with as far as leveraging AI within our businesses?

3:42Yeah, okay. First of all, I think you should test it by yourself. That's obviously the most important thing. And the good news is that it has become very easy. GPD is not that young. It has been released in 2020 actually, but it became really popular last December with chat.jpg where everyone was able to do some tests. And now there has been some extensions such as chat.jpg plugins. So you're becoming more and more aware of its abilities, what would work and what would not work. If you want to like having some rule of thumb of how to decide if it is a relevant use case or not, I would say is, can you put an intern on this task and would it be working?

4:29Like, for example, if you are asking an intern to do some research on the discussion of a customer inside a spreadsheet, okay, if it works with an intern, it's sure that it's going to work with a GPT. But the opposite, so if you want to ask your intern how to do a business plan for your next company, you in turn will do something that might look like a business plan. But if you go deeply into it, you will see that, okay, this is mostly garbage. So that's where we are right now in artificial intelligence, which is quite amazing because a couple of years ago, we were just at the level of babies.

5:05And now we are achieving the level of interns. I guess maybe in a couple of years, we will get the level of developers. And so far, I think our jobs are safe, but we should be careful about it is not sure. I love what you're saying. I think addressing AI as an intern is probably the perfect, easiest way I heard anybody explain on how to use it in business. I think the interesting is you can take it to the next level, meaning to maybe not the CEO, but the next level and what it can do above an intern if you A, give it access to the right data within the organization and B, have somebody monitoring what it's actually doing.

5:46So if you're giving it access to more stuff, like to your CRM or to your FAQs or to whatever data you have in your company, history of emails, whatever data you give it, it will be able to do more than an intern as long as you're actually monitoring what it's doing because it will sometimes do really weird things unexpectedly. and the crazy thing will do it after it's quote unquote earned your trust because we'll do it perfectly fine 76 times and then in the 77 time we'll do something really stupid totally unexpected and if you're not monitoring it 100 % of the time every now and then it will surprise you and that surprise may cost you a lot of money for a business that's the only thing to know and that I would add to that so now you're saying okay it's an intern and you understand the tasks you give interns are usually very well-defined, repetitive, clear tasks.

6:41That's what you mean? Exactly. So one of the, you said very quickly is that you need to be very specific on the task. On an intern, you don't want just to say, okay, makes money for me. It's not how it works. You want to see, to give him some context about, okay, you are going to represent a company. These are the values of our companies and you want to be an ambassador for us. You are going to talk with like this. And what is also interesting is to providing some example, what type of message can you say? What can you not say? Actually the first, so the paper GPT-3 is called a language model of few short learners, which means that they can learn just from a few examples, what they are supposed to do.

7:29and this is a really amazing feature to say okay i'm just providing you two examples and okay now you know what i'm supposed to be doing it's the same with an intern actually if you want an intern to work correctly you have if you're just providing some really broad context and not monitoring what he's doing it might fail quite often but if you're helping it to see what is the quality of his job and then you're interacting you're repeating you and improving your indications and your intern is going to do a better job. Yeah, I think that's some rule of thumb of what you can do. Okay, so we're saying you want to use AI for stuff that's very clearly defined.

8:09You want to be able to have some examples of what to do and not to do. How do we take that and translate it into the actual world? And again, the framework you showed me blew my mind. And I'm going to tell everybody and frame it this way. It's a visual process. Think about a flow chart that you create. And I'm going to share this on social media, and on other channels and YouTube potentially, so you can see this. But we're going to explain this step by step in a way that you can visualize it in your head. And we're not going to create a crazy complex process that will be relatively easy to follow.

8:43But Pierre, let's really start with a use case that you want to define. And let's define the steps of it and how it can be built. And really the amazing thing, it's free and it doesn't require any code and it creates amazing results. Yeah, one of the things that's how you can think about GPD is to say that this is a glue between your existing software and your customers, like your ambassador that is representing your trade. So if we want to use one use case to show off this feature, this benefit, what we could say, let's imagine a case where we are sending some message to our prospects, to LinkedIn, to email or whatever communication tool.

9:27And then we want to see, okay, this potential customer is answering to our first message. And we want to detect first if this customer sounds interesting or not about our introduction message. And second step, we want to send him an answer to extend the discussion. We could make it more complex by then connecting it concretely to our customer relationship management software. or also maybe trying to decide when it is relevant to have a human in the loop. But I guess so far, it's a good idea to keep it simple and stupid and see what we can do in just a couple of minutes. Okay, so let's do it. So what's step one?

10:13Okay, so step one here in our case, we want to see that that's when the customer is answering to our first message. So we want to... I'll pause you for one thing. Let's mention the tool you're doing this in, because again, if somebody finds this tool because of this show, it was worth listening to the show. Yeah, okay, okay. So here we're going to use NodeMation. NodeMation is an amazing tool that has a huge community inside around it. It's open source and you can think about it as Zapier, but as an open source product. Open source means that it can be free of charge. It can also solve some more complex cases.

10:55It might come to the cause of being slightly more complex for first user, but I think it's totally doable even for first user. So it's called not mission, but most of the time you will see it's written as N8N. It's a geek thing of saying what is the first letter, the last letter, and the number of letters inside. It's something we do usually. So here, that's the name of this tool. Okay. Okay. And let's see how we can solve this use case of automatizing prospect, you know, but defining what is the first step. So the first step is when a customer is sending us an email and we want first to receive this email.

11:41So what we are going to, we are going to create a node into our pipeline. So I have here a big plus. I can click on it and saying that I want to have this special note that is triggered as soon as I'm receiving an email. So once I have it, I can... Just to pause you for a second, for people who use Zapier and any other automation tool, you're basically telling it, okay, start this process when an email comes into this email address. That's basically all it's doing. And then it can capture all the data from that email once you connect your email into this platform. Yeah, exactly. So you can really have all the data.

12:21That's why it's a bit more complex than Zapier, because it really lets you decide what type of data is important or not. So you can see it in a schema format. And so now we can test it and wait for an event. So I can send an email to myself on which I'm going to write, OK, your product looks really cool. I want to know more about it. And now I can see this email in my workflow with all the other metadata. So what else can I do? So now I have my message and I want GPT to decide if this message sounds like it is, the user is interested or not. So I'm going to create a single node that is going to be connected to the first one, an open AI node.

13:10And on it, I'm going to drag and drop my message. and also create a prompt. On the prompt, I will say to GPT, okay, the user has sent this message to my column and then I'm putting the message of the user. Do you think that the user is more likely to be interested or not about our product? Just answer true or false without any explanation. So what we are doing here is some prompt engineering. So if you're a bit familiar with GPT, you might know that sometimes it tends to get confused easily. So you want to provide very specific information and avoiding to fall in some sort of trap. You're going to put also an option about the temperature.

13:58So here we can say... So before you dive into temperature, again, just to explain to people, in step one, we said, okay, did I receive an email? Okay, if I receive an email, send the content of the email and tell it to look at what the email says and decide whether it's more likely or less likely. It's not a hundred or a zero. More likely or less likely, that person that wrote that email is interested in the product that we offered him in the original message. And we asked that GPT to answer in true or false because we're going to use that as the trigger for the next step. So just to put everything together and what we've done in these two steps.

14:36And now temperature, I'll let you and explain what that is as well for people who do not know. Yeah, so we are going to add an option with the temperature. We're going to set it at zero. Why? Because temperature is reflecting like how random chat GPT is. So in some cases, you want to be a bit more creative. Like for example, the product was creating some lesson to learn Italian. I want it to be creative here. I just wanted to classify if my user sounds interested or not. So I don't want any surprise. I don't want any randomness and I'm going to set the temperature to zero. So when I'm executing these nodes, I can see that GPT has answered true.

15:21So it means that it has returned literally true. So he thinks that my user is interested about my product. It's nice because like somebody said, it was pretty clear. The product looks really cool. I want to know more about it. So indeed, that's what I wanted it to do. In case that GPT is not working so well, what you can do is also on your prompt, you're providing to GPT, adding some example. So, for example, in my prompt message, I could write, okay, example of message, your product is really bad, answer, false. And second example, okay, I'm not sure if your product is fine or if it really fits my use case or not.

16:01in this case, you want to send a follower message. So you want GPT to decide that it is true. And it sounds obvious for you, but thinks again, that it is just one day intern that doesn't know much about your product and business. So you want to be very specific by providing some useful example. Yeah. To add to this, you can use historical data, right? You can take actual examples of messages that you've got and what an actual intern or the person who does this in your marketing or sales team has done and say, in this case, it's true. In this case, and literally copy and paste, that's option number one, if you want to copy and paste.

16:41Option number two, if you have this in some kind of a Google spreadsheet, you can just reference that spreadsheet with a true or false as examples, which would then feed into the calculation of what ChatGPT will answer. But like Pierre-Louis said in the beginning, just give it examples. And the more examples you're going to give it, the more accurate to the way you want it to be, ChatGPT will actually give you the answers. Yeah, exactly. Exactly. So think about it as an iterative process. So you do some attempts, then you see, you look at the outputs, you monitor it, and then you decide, okay, I want to improve there and there because sometimes it's doing not exactly the correct job.

17:22So sometimes you should not expect it to work perfectly fine at beginning. It's not a magical tool, but you want to iterate on it and doing more and more prompt engineering. And in the end, you can get really, really good performance. Awesome. So at the end of this step, ChatGPT is going to spit out an answer, true or false, whether that user is interested or not interested, most likely interested or most likely not interested in our product based on the message that they've emailed us. Yeah. Yeah. Thanks. So next step is to add a condition. We want to have different branch inside our workflow.

17:58One branch, if the guy is interested and another one when he's not interested. So I'm adding another node, which is called just an if, and I can put a collision on this node. What I'm going to do here is that I'm going to drag and drop the answer from open AI. So it has answer true. And I'm going to put a condition saying, okay, does this value contains true? If it is the case, then my message will go to the true branch. And if it is not the case, it's going to go obviously to the fourth branch. So once we are at this stage, so we can see that in our example, the message that we send to an email has arrived on the true branch of this condition.

18:46Okay, great. And now we can use create another node with GPT to write a message, an answering message. So what we're going to do is selecting our original message, the one that the user sent inside the email. And again, I can drag and drop inside my prompt. And on my phone, I can ask GPT to write an answer saying, okay, for example, schedule a meeting with a user. Let's see that maybe you can connect it with your calendar to see your availabilities. And you can list like five different options that is working fine for you. And you want GPT to summarize it into an engaging and with message. So again, this goes back to all we've done so far is we got an email.

19:35We said, is this email positive or negative with regards to an interest in our product? We know that in this case, it's positive because the user literally wrote to us, your product is really cool. I love it. So we know he wants to do. And now what we're doing, we're using ChatGPT to create a personalized message that is using the original message. So we know what that person actually has written to us. and we're writing a prompt that will help us set a meeting with a person using whatever external tool that we have that we're using, like Calendly. And so all these steps are happening automatically if you write the right prompt.

20:18So if you can share what kind of prompt you would use, that would be helpful, I think. Sure, sure. So here, let's do the test. It's a live demo. It's never a good idea to do a live demo, but give it a try. So we write as a message, as a prompt. So here is the message of the user. So here we have drag and drop this message and then write a personal message to find a meeting with him. Availabilities are Monday from 5 to 6 p.m. Tuesday from 2 to 6 p.m. What I want to do then is setting also another condition, which is another option, which is the maximum number of tokens. Let's put it to a very high number, like 1000, for example.

21:04So this is limiting the length of the message generated by OpenAI. So the trade-off here for anybody who doesn't understand tokens, tokens is basically the amount of data that you can get out of it. And you're paying for that data. This is what you're actually paying for when you're paying for ChatGPT, you're paying for tokens. And the higher the number, the more it can write, but the more you're going to pay. There's also, not to confuse anybody, but when you use the API, there's different levels of the API you can choose. The more advanced, the more you're going to pay. So there's a trade-off always in depending on the tasks, how many tokens you want to use and how advanced of a model you want to use to do these kinds of things.

21:46And you can learn that trade-off either by just doing your research or by testing stuff You can try different levels of the model and see if they're good enough. Awesome. They're going to work faster and they're going to cost you less money. And if you want to write really sophisticated, longer stuff, like a full blog post on a topic, then you probably want to pay for the more advanced model. Exactly. Yes. So here is the answer from the chatbot. He said, hi there. Thank you for your interest in our product. I will be happy to share more information about it with you. When are you available for meeting?

22:23My availability for the upcoming week is Monday from 5 p.m. to 6 p.m. and Tuesday from 2 p.m. to 6 p.m. Please let me know if any of this time works for you. Thanks! So here you might want to monitor it very carefully because you want to inject some information, some key information, such as your availabilities. So if you make a mistake on it, it's not a big deal because then you can just say, okay, actually I was not available, but then you can also use some technique to avoid that OpenAI or GPT is missing with your message and hallucinating dates that you're not available on. For example, you can write some special tags saying, okay, are you sure that this information is really contained on the original message?

23:10And so this type of hack is reducing significantly the risk of making some mistakes. Quick question that I think I know the answer for. I can use my actual Calendly link. I said, I want you to use this link within the message that you're writing as my availability and connect my Calendly to that and it will actually do it, right? Yeah, exactly. So what you can add on an automation, there is a special note for Calendly that you could use as a trigger or just to observe when are your available spots. So, for example, you could write another workflow saying that each time that someone has booked a message, has booked a spot on my Calendly, I want to send him like a message, like a reminder one hour ago, one hour before, and so on.

24:02That's pretty easy to do. personalized messages that's really cool yeah okay so going back to our flow now we have written a response message to the person that has sent us the email that we've identified as somebody who's interested we've written him a personalized response that actually takes into account what he has written to us and our availability what's the next step so we are almost done we just need now to send the email to to his email address so here we have imagine that our use case is that he's sending an email so we want to send him an email back. Actually we might want to put some waiting time because it sounds a bit suspicious if I'm receiving it like and so one minute after I could add a special block for slipping the cards for a random amount of time let's say half an hour one hour or I could just directly send him the message so So here is going to be very easy.

24:59I'm just going to drag and drop my message on the text. So the message that was created by Chachapiti dragged into the step of the process. Yeah, exactly. I'm going to drag and drop also the people who sent like the email, his own email from the original email with which he sent me a message. So I'm going to put this reference in the two fields. and I'm going to say that the form will be from my email. I guess this is like, you know. So again, basically what we're doing in this step is we're taking the email address of the person that sent us the email. We're taking the message from ChatGPT and we're putting them together in an automation step that will send them the email from my email address that is connected to this tool to them because that's where they sent the email from with the message that ChatGPT created specifically for that user.

25:54so we send it and now we should see the answer in a couple of minutes once like it is put inside the queue we had a positive message from MedGun which is the service I'm using to send email and in a couple of minutes I should receive this email back into my email address so I want to summarize this particular use case and then I want to generalize it with you for additional use cases. What this tool does, so what NodeNation does, is it really allows you to build a step-by-step process for more or less anything you want. But the beauty now is while Zapier, you can create all these steps and connect all these tools.

26:33Now you can connect tools which have OpenAI, ChatGPT built into them, which allows you to do things that were not available before, such as in this particular example, classification of things. So which bucket it falls into. So you You can then define different branches of what's going to happen in the process, as well as personalizing anything you want based on whatever data was there before, because it can write based on the previous data that existed. And that data can come from communication. So any communication channel that you had, like on LinkedIn or messaging or emails, or from your own internal data, like CRM, if you connect it in there and bring it CRM data.

27:11And then the outcome can be still whatever you want. It can send a message back. It can write an email. It can change fields in the CRM. It can do whatever you want it to do. And all of that happens very personalized while analyzing nuances that you can train it to do. You can create these processes in minutes. We just did it together while walking people through it without writing any code. And I find this literally mind-blowing. I want you to give a few other examples that you use this for, even just as tests or for real of use cases on how to use this tool for stuff that people can just have ideas that they can play with yeah sure so i have a couple of them here maybe i can go so i have another one that i've never talked about it with you before let's see i will show you a demo so i created a bot a robot on telegram so telegram is an instant messaging application and yeah so this is so my partner she's an english teacher and she so when you're a teacher you have to correct tons and tons of of written exams yeah and so here during our vacation she was spending all her days doing this correction i was okay i don't want you to spend all of time doing that so can i help you somehow to help you to correct all your exams.

28:39Obviously, I'm not as good as she is in English. I had to use other things I could do. And what I did is that I took a picture of the subject of this test and I put it in an API that is doing some handwritten text recognition to have a written example, like a text message. So basically taking handwritten exam and translating it into a digital text that can appear in a message or an email or anything anywhere else. Exactly. Yes. And then I asked GPT, OK, can you correct it and answer back the message on Telegram? And I had to put, again, some prompt engineering. So what I wanted to do is that I want first to classify, to detect all the grammar and vocabulary errors.

29:26So for example, it would say, okay, oppose instead of opposed in sentence one of exercise one and so on. So for example, defense meritocracy, okay, this one is maybe like influential instead of influence in sentence four of exercise one. And also I ask it to detect some incorrect facts or unclear statements. So for example, you could see that some dates, some historical dates are incorrect. So obviously it's not going to work all of the time, but it's already a very promising proof of concept. So again, just to talk about use cases, in this case, you're taking an image of handwritten stuff, analyzing the data, providing feedback from the data, and then you can decide what you want to do with it.