Transcription
Um, yes, welcome to the uh next practical webinar from our series. We've been doing this once a month now. Um, an introduction to the possibilities of AI automation for your own business processes. And this is something that you, um, see very, very often on social media, uh, like LinkedIn, maybe also Instagram, etc., that, um, agencies tell you, yes, we can automate this for you. That's usually not real estate specific, and I think it's always good to let those automate things who, um, have a clue about the topic, have a clue about real estate. Um, nevertheless, these are of course, yes, uh, these are of course, um, good offers, good things, because, uh, you can automate and digitize a lot via AI, etc. And, um, therefore, if you are interested in this, make sure you either do it yourself. I will, uh, give you the first hints today, of course, and, uh, and also show you various steps, um, on how to do this. Um, on the other hand, there are agencies for this now. Uh, I will also comment on that again in the coming, uh, coming days on LinkedIn. Um, uh, you can also take a look at it. Um, but a practice-oriented introduction, we'll dive directly into this topic. But before that, I would like to tell you very briefly why, uh, I am doing this with you today, and accompanying you. Um, I have been part of the real estate industry for 7 or 8 years, was responsible for digital strategy at BNP Paribas as a project manager. I am not a developer, I am not a techie. Um, I, uh, I am perhaps like many of you, uh, I inwardly run away when I see code because I think, no, that's like a new language, I can't learn that now, I can't teach myself that. But, uh, AI can do it, so AI knows it and AI can also tell me, uh, where to insert this code and how, in order, for example, yes, to build things, to build small tools, to automate something. Um, that is one possibility. The other possibility is always with, uh, with tools like Makcom, SPIAN. Um, these are all tools that we will look at today, or SPIA and NN at least. Um, to build your first own automations with them, to look at yourself how things work. And that is also all possible without really knowing coding and without knowing how to connect programs and tools at all. And all this is only possible because for 3 years now, I think about 3 years and four days, TBT has existed. Uh, for 3 years, one month, and four days, I have been working at Site. That has also, uh, turned out quite well, I would say. But these are all things where you can build these automations for yourself with the help of AI, and, um, honestly, even if you had no idea before, you will see the first successes after a very short time. And some of you will be so interested that you say, okay, I'll take care of that more closely, and build it up a bit further, because you quickly have a sense of accomplishment with it. For others, it's like, okay, nice, cool, I see it, I know it, I understand it, but someone else should build it. Um, and, uh, yes, for both, for both sides, uh, of the coin, this webinar is very, very, uh, exciting, and, uh, simply seeing what is possible and how it all works is, I believe, the basis for being able to assess well what you actually need or what you can actually do in your own work. Um, and on these topics of automation in general, artificial intelligence, I have been, uh, for about one and a half to two years now, and I've been saying this for so long that I think it's two years, I would have to check again, I am now a lecturer, uh, at ADI, at CIA BDB, at IRAPS, and I give workshops in companies, I give workshops, and lectures, uh, for other organizations, etc. Um, so I do what we are doing here together today, uh, very regularly, but every time I must also say, every time it is super individual for the, for the own, uh, case, because, uh, I don't choose that myself. Uh, I would of course like to have a presentation that I can copy every time, but, uh, that's not possible with AI, honestly. Everything is changing so quickly. Uh, within two weeks, what I built two weeks ago is, uh, somehow obsolete in 5 hours, because after two weeks, a new tool comes out and then you see, oops, uh, it only takes half an hour now, and, uh, it's really something where you have to stay a little bit on top of it all the time to, uh, to really, really get into it. In the beginning, it was like this, uh, with these presentations, I, I must honestly say, I no longer felt like spending four hours, uh, to then build a presentation, and I said, I'll make the presentation as simple as possible. Um, at that time, I didn't use any AI tools for it. Nowadays, it's Gamma, with which I can build the layout of the presentations. Uh, that's of course much faster. Uh, but back then, I said to myself, hey, yes, we'll just go directly into the tools with a practical approach, and I'll really show you what's possible with them and how it works. And then at some point I realized, I would like to learn like that too. And it was also well received. That's why, um, we're doing it today. We're really going into the tools, building things together, and, uh, looking at how it actually works. I can always show you or tell you a bit about the pitfalls, uh, that might exist, uh, where you might have to work particularly long, etc. That means real tools, real application examples from the real estate world. Um, that's what we'll look at today. And, um, maybe you've seen Site in one or the other, uh, presentation, workshop, webinar that I, uh, that we are doing here. Um, I don't think it will change very quickly. Um, just to clarify the expectations, AI is not a magic bullet. Yes, it's not the tool, uh, to which you say, I am a real estate agent, do my job for me, or I am a project developer, do my job for me. Um, and then it does everything automatically. No, that's not it, but honestly, not yet, because, um, everything we do on the PC, on the computer, on the, uh, on the screen, let's say, especially with data, compiling things, extracting data, researching data, etc., these are all things that AI can already do. Um, but we just have to think about these processes that we see in front of us, that we actually want to have done by AI, uh, to think, okay, um, what are the individual small sub-processes? And now think of any process from your, uh, daily work, what I always like to use is, uh, project developer, existing property owner, receives an exposé, receives an email with an exposé, and then it's checked, does it fit my acquisition profile or not? And just reading this, reading the exposé, then comparing it with our, uh, with our acquisition profile. That's something that AI can do perfectly. Yes. Uh, and there are also, if you prompt it properly and, uh, know what it's about, there are honestly fewer, uh, fewer points where you, um, I think you can hear the ringing here, if I get an email. Of course, I'll leave it for now, uh, if it gets too much, I, um, and this, uh, this comparison of the exposé and the acquisition profile, maybe they'll tell you that the AI could be uncertain, could be wrong, could give out wrong information, this so-called hallucination, as it's called. Um, if you prompt properly and set it up properly, then you have this with such a low probability, uh, it's actually heading towards 0%. And, um, honestly, you can also overlook something. Yes, so we always want the AI to be 100% perfect, but humans aren't either. And AI is much better and much more accurate if you do it right. And in this process, acquisition profile, email comes in, etc., all the small individual sub-processes. Email comes in, first to see what kind of email it is. Do I have an attachment or exposé, or do I get it via a link or something? AI can do that, AI sees that immediately. Yes, downloading the exposé, and, uh, extracting the data from it can also be done by AI. Um, then comparing it with the acquisition profile can also be done by AI. Then to say, okay, we are still missing this and that and that. Um, I have to reply to the email now and, uh, ask the broker for it again. For example, AI can also do that, and if you keep thinking in individual small process steps like this, um, you will find that we can actually have AI do everything we do on the PC. And that is then, um, the possibility to build this magic bullet for ourselves. So in the very distant future, um, yes, with a lot of work involved, of course, because, no, I am a real estate agent, do my job. Uh, it's not just these processes, but a lot, a lot more, where you always have to think, uh, yes, no, set some filters, uh, what, what is the next step, etc. But, um, that's everything that can be automated, that works, that you can do. You just have to take the time for it. AI will, AI will always get better. Uh, especially the Large Language Models. It's actually honestly Large Language Models that we are primarily talking about here. They are getting better and better, and therefore, uh, it's getting closer and closer that we can actually build it like that. And the key to that are then, yes, I've already mentioned what's there. Key to success: Large Language Models. Yes. Uh, and the key to actually building it are automation tools. That's why here are Vapier, Makcom, but especially Natn, with which we can build it. There are other tools. Relevance is called that, for example, or a tool is called that, for example, um, it's already at the point where you just enter in language, i.e., text, what you want, what this automation should be able to do, and what it should think and consider, in which cases, no, it does that. Um, and you enter that there, and then you get a final, finished process with individual steps next to each other, one after the other, which you can execute. I don't have Relevance with me today, or not with me. Um, in my tests, it's usually the case that when it gets a bit more complex, uh, you have to adapt a lot yourself. Um, Relevance is not necessarily the perfect tool for a beginner, um, because you really have to go into the code a bit, but AI also helps in case of doubt. We'll see that in a moment when we look at End. Um, but such things exist, and they all work. Yes, so, um, I think there is a strong business opportunity if you say, I want to automate everything here for me, and take the time to do it, or find the experts who will do it with me. Um, you have super many possibilities with it. Yes, um, the complexity ladder, we've called it that, and by "we" I mean, of course, AI and me. Yes. Um, what, what have I thought about what we will do today? Uh, we will actually go into four different areas of automation and, uh, see what is already an automation, also with Custom GPTs from ChatGPT, for example, what can we automate with it in our daily work? Um, then the next step is that we take Langdog. That's something like ChatGPT, I'll go into that in more detail in a moment, and bring that together with integrations. Integrations means, uh, we take our Outlook, for me it's Gmail, account in there, so it can access emails, for example, and, uh, I'll show you everything in a moment. Then the next step is, we'll take a quick look at, uh, SIA. With SPIA, you can see quite well how an automation works. First this happens, then this happens, then this happens. Um, and that was here, and there you can very easily, even as a beginner, build such an automation. And as a fourth step, we'll go to NN. Uh, that's also relatively easy. It looks a bit more complex than Selpia at first glance, but honestly, if you really want to, uh, deal intensively with this topic, then you should, uh, definitely learn Nat and, uh, look at how it all works. But we're making it complex, so it gets slightly more complex each time, but even the last step isn't really complex. It's only a few, uh, so you'll probably have built your first very simple automation with NN, if you've never used it fully, in, I'd say, 2 hours. Um, because you simply know that you can ask AI to support you in every way. I'll show you exactly how that works in a moment, too. Uh, because that's how I honestly built the example. Yes. Um, Topic 1, lowest complexity, a Custom GPT from ChatGPT. What does it do? Um, it basically does what you tell it to do, over and over again. And for that, we'll go to ChatGPT. I'll show you, um, what it does, using the example of an, uh, email creator. In this email creator, I have, uh, I'll show you here. So, if you want to create a GPT, click on Explore, then you'll see Create in the top right, and there you create a so-called Custom GPT. Um, I'll go into edit mode now, and, uh, then it will look exactly like this. Here you enter instructions. This is basically the prompt, uh, what the GPT should do for you repeatedly. With this example, uh, it now creates my emails, and I basically just have to enter a prompt on how I want these emails to be created. I tell it here, um, you will receive, uh, loose keywords, you will receive information in dictation style, uh, on how to write this email, you might get incomplete sentences, and, and then, um, it might also be that I copy an existing email and then say, reply to it with this and that content, um, you have to tell the AI this. Then, uh, it should write professionally, no promotional tone, no emojis, uh, etc. It should include links, and, and, um, we won't go through it in detail here. That's, that's not the goal at all, but the goal is that you see, you enter a, um, system prompt, as it would be called, once, and then you have it in there. Now I'll go back to the GPT here, and now the GPT always knows what to do when I enter something here. If I say, for example, uh, email to Mr. Müller. Thank you for the friendly conversation here, the presentation of Side, and here, um, my appointment booking link for the next exchange. Yes, by the way, I've forgotten how to type. I used to place a lot of importance on writing everything correctly. Nowadays, not because AI recognizes everything. It doesn't matter at all. You get the feeling you just hit the keyboard, uh, and then, uh, and then the email comes out accordingly, because it understands everything. Yes, now you see here, um, it knows exactly what to do with this. Yes, well, I also said email to Mr. Müller. Um, but still, it knows immediately what to do. It knows that I always write "Side" and a dash in the subject line, and then it has put the text here, and these are links for appointment booking, for example. Um, I provided the link beforehand, uh, so it can immediately do something with it. And now I could copy and forward it. And the GPT or copy and write the email. Yes, and the GPT is now basically just the statement, um, whenever this happens, please do that. And that's exactly what, uh, we'll do now by building this GPT, exploring and creating. And we'll take an example. I've already prepared a prompt for this. I'll copy it in there, and we'll go through it very briefly. Um, no, I'll copy it in here, and we'll call it Exposé Evaluator. And we don't need to click anything else right now. Uh, a description is usually irrelevant. It doesn't help the GPT to work. Voice prompts either. You could store knowledge here, for example, acquisition profile, I haven't done that in this case. What is important here, in this specific case, is that you enable Code Interpreter and Data Analysis. Why? Um, well, what we will upload here, namely an exposé that should then be output in a structured way. Um, this exposé can be available as a Doc file, as a PDF, or as a PowerPoint, or something similar. And these files, um, they are not automatically understood by the AI. What, uh, ChatGPT, what other AI does, is basically, they build a tool for themselves, uh, with a script, or they build a script, or they build a tool with which they can extract the information from this file. That means it's always very important to enable this, and ideally, also to say in the prompt, you, um, if you get a PDF document, then extract the data from it with a script. Yes, then it builds this script itself, or with an OCR, which is, uh, the recognition of images or the recognition of text in images or in PDFs, for example, um, that would also be something that can be used, and the Large Language Model works with that accordingly. Yes. Um, and in this prompt now, um, it's very, very extensive. I'll come to how to create these prompts in a moment, because I wouldn't want to create them myself out of laziness. In this prompt, we are now simply saying, role and goal. You are an AI assistant for Pro Wohnen Projektentwicklung GmbH. I've invented a company there. Um, read emails plus exposés, extract all acquisition-relevant data, and output it in a structured JSON format, compare the offer with a fixed acquisition profile. And what happens here is, uh, first the data is extracted in a JSON file or in JSON format. We don't need to go into detail about what that is, but it looks like this. It's actually irrelevant for our use. So, if the AI gives you the prompt output, makes JSON, then you simply transfer the prompt, put it in there, and that's it. Done. Yes, so you don't necessarily have to understand it in this case. Um, then a short comparison with the acquisition profile, a short, uniform assessment in, uh, budget points. And this acquisition profile that we have, uh, can be found here. Yes, this is the short version of the acquisition profile. I've also made it completely fictional. Um, and basically, what happens now is. We get, we copy the email from the broker plus the exposé in here. It knows how to, uh, how to work with it. And what it then does is extract the data from the exposé, uh, output it to me in a structured way, and compare it with the acquisition profile that we have. Does it fit or not? And if we need data in the acquisition profile that is not available in the exposé, then it automatically creates the, uh, reply to the email, which I then have to copy and enter, or copy and paste, and to send an email. Then it automatically creates this reply to the email with the statement: "Look, uh, thank you dear broker, um, I still need this and that and that from you, so that I can properly check it." That's basically what's in this prompt. Um, just briefly on the topic, I didn't write the prompt myself. Yes, the AI wrote it for me. What I would definitely recommend to you is the topic of metaprompting. We can, um, we can, uh, with a few instructions, what the GPT should do or what a prompt should do, very easily and very, very quickly, create such longer prompts. I'll show you that very briefly. Um, what I use here is this prompt enhancer. That's also a GPT. In this GPT, it simply says, look, I'll tell you what I want, and you use, um, the prompting skills and tips and research the, uh, possibilities for prompting, and give me your perfect prompt. That means I just tell it, um, we can do it very briefly now, not for this, for this example. Um, I need the perfect prompt for a Custom GPT with which I, uh, when uploading an exposé for real estate, receive a structured evaluation of the data directly. So, that's really very simple now, and the prompt won't be as long as the one I have before. But you see, I'm getting out, you are a professional real estate analyst and data-driven exposé evaluator. These are prompting things that must be given, rules that make the prompt better in the end and make the AI understand it better. Rules: Read all available content. All results must be traceable and cleanly structured. Output format: this, please. So, it has already gone a bit into the content and is now giving me this very long prompt. Um, the one we'll use in a moment is even longer, because I naturally told it more about what it should do. But just let the AI generate the prompt for you, and also let the AI generate the prompt for, uh, creating such a prompt enhancer, and then you can use it like this at any time. And these are much better prompts than what we would enter, because when we have time, we might not think of everything and, and. Yes, and that's how this prompt came about, and we'll use it. I mean, the prompt came about because I also uploaded my acquisition profile, my fictional acquisition profile, and said, um, now extract the most important data from it and, uh, incorporate it into this GPT. Now we'll create this GPT. Yes, anyone can use it with the link. That's also completely, uh, completely irrelevant. Um, what I want to show you briefly in the meantime is the exposé, which we'll use as an example in a moment. Ah, okay, too bad. Um, I can't show you that right now. I didn't consider one thing. That means I'll do it this way. I'll open it individually. Just so you can see that it's not some nonsense that's coming in, uh, and then you can say, yes, it only works because something else came in. Um, I'll share with you, where is it? Pages, so share. Share. Now you should see this. Yes, this is the, um, completely fictional exposé for a property, uh, which I, uh, created for myself. Yes. A bit too much information, perhaps, depending on it. Yes. Uh, we have a raised hand. Is that, uh, a question whether you can see something here? Uh, so if you, if you don't see something, then, um, because it's not working somehow. I actually think it's not working here. Okay, share now. Okay, next time, if you should not see something, raise your hand or or write in the Q&A, uh, in the questions and answers section, please. Uh, then I'll see it immediately. Yes, this is the exposé that we'll drag and upload. Um, I also have a, uh, a sample acquisition profile. Um, which we'll compare it against, I can show you that too, but, um, maybe we'll do that in a moment. It's probably not that relevant right now, uh, because we already have the acquisition profile in the, um, in the prompt. So. Uh, now we go back to ChatGPT. By the way, I just saw in the questions and answers, there's a question about whether you'll get the prompts later. Um, generally, we don't do that, honestly. Um, we'll send you the recording, uh, of this, but, um, the, yes, slides, not really. I'll have to see, maybe we'll do it a bit differently this time. Um, and then I can also include the prompts. That's actually not that much effort. Uh, but we get this question quite often. Yes. Um, now we go back to ChatGPT, to the, uh, to the exposé evaluator that we just created. And I'll go back, because the exposé evaluator is found here on the left side. We click on it. And, uh, now we take the, which is why my emails were open the whole time. Um, I received this email, so it's completely, uh, fictional again. I received this email, and because the, uh, GPT is waiting for me to enter an email here, and at the same time this, um, exposé, which we still have here, so what I just showed you, uh. And, uh, therefore, we'll go back to the GPT with this and upload the exposé here as well. And now it should recognize that this is an email, and it should recognize that this is an exposé, and then do exactly what we just told it to do. And that's the first, yes, a first partial automation in the end, because we, uh, have a process that we would normally do, namely extracting the data from the exposé. You see, GZ GFZ is displayed as you saw it before. Um, this is something that we might otherwise do manually or enter into some lists. In this case, we get it out like this first, and also get this short comparison with the acquisition profile directly. Yes, Hilden is clearly in the core area. Plot size is optimal within the target range. GZ GFZ and BGF are exactly within the target corridor. Uh, B requirement met, perfect match, and, and now the question is, which documents are still missing from what we actually see in our acquisition profile? And the documents that are missing here are the development plan extract, the site plan, the land register extract, and, and, and with some information it's unclear. For example, we don't have a concrete purchase price expectation in there at all. Yes, and that's why it has now created this email for us. Dear Mr. Wagner, this is the fictional broker for sending the exposé, the friendly telephone coordination. Um, yes, uh, it was mentioned in the email that we had already spoken on the phone about this, so it's included here. Um, we still need the following documents. Specifically, we need exactly this. Yes, as soon as we have these documents, we can confidentially, uh, process the exposé. Uh, we can confidentially, uh, review it. And now I could just, if I had entered the name beforehand, it would be even better. Um, and now I can copy this, uh, and reply to the email, and would then reply to the broker accordingly. We have a question about this. If a property is within the scope of a development plan, can that also be evaluated? First, we are currently performing the, uh, evaluation of what is in the exposé. Um, if it's stated in the exposé like that, then, uh, it will naturally be output here during the extraction of this data. This is not further research, etc., at the moment. Um, that means it would display it here. Um, but you can also find out whether this property is in a, uh, development plan area or not, with, uh, AI tools like ChatGPT, Deep Research, and the ChatGPT agent mode. Uh, these are then automations that you would then build around it step by step. Yes. Um, but that is certainly possible and it works too. Yes. Um, exactly. But this is already the first, or the first partial automation of what we can do here. This is step one. This is actually still the, uh, the easiest and simplest thing that you can build yourself within a few minutes, even if you're doing it for the first time, in the end. Yes. Um, I'll close this again so we don't hear this email. Um, we don't need it anymore, this email at least, uh, because we'll now go directly into, uh, the second, uh, part of the automation, and for that, we'll use LDOG in a moment. But before we use LDOG, a very brief excursion and detour, uh, to so-called APIs, to so-called interfaces. Maybe you've, uh, come across this term somewhere, heard it. Maybe you're already working with it. That's also possible, um, APIs, yes, Application Programming Interface, as I've learned myself after using it for years. Um, APIs are, uh, basically the connection of two programs to each other. Yes, what we did earlier, letting ChatGPT create the email and then going into Gmail to send that email. That would be something that can be completely automated via APIs, via interfaces. So, this process we just had can be automated if you let the programs talk to each other directly. The vast majority of the tools that you use daily, whether it's, uh, Onffice for brokers or a CRM tool like HubSpot or something. Um, the vast majority of tools have such an API.
API. This means that another program besides this actual tool can automatically access the tool with the API, make changes within it, add data, retrieve data, and work with it further. And there is no, yes, intermediary in between. Um, we will build this intermediary shortly. Uh, so it would be possible for them to speak to each other. We will build a few intermediaries shortly, uh, so that we can build it ourselves and don't have to be the biggest IT ITers and IT people for it. Yes. But these programs can all speak to each other, and this really applies to all programs that you have. So, imagine if you receive an email and, uh, there, um, you are sent this exposé, then you can immediately have it sent to ChatGPT. It should be extracted, and then the extracted data should be immediately stored in your tool where you collect exposés and information about properties. That is possible. They work with each other, among themselves, and are very, very easy to connect with each other. I believe that has something, um, I believe that was a somewhat fitting explanation, uh, for what is happening here, because, uh, we will now, uh, we will now actively use it shortly. Perhaps, before that, on the topic of data protection, if you build a tool yourself that does this, or if you build this connection of two tools yourself, then only these tools work with each other. If you are already working with the tools anyway, then, uh, you actually have no data protection issue that you might not have had before, because, uh, in the end, if you have checked both tools for data protection issues beforehand, then connecting the two tools is also not a problem. Yes. Um, if you use automation tools like Langdog, Make, NLN, uh, these are all tools where you take an intermediary in there. But you only take this intermediary in there, uh, to build it yourself. So, it will then be sent to the other tool, for example, via Zapier. You actually do this because you can build it together more individually, faster, and significantly better, and build it together to say, this is what's actually happening here. Um, tools like NN. NN is, for example, a German tool. Um, it is completely, uh, uh, completely easy to use and compliant from a data protection perspective. Um, NNN also offers the option, uh, to host the whole thing on your own server, uh, so that nothing really works outside of your spectrum anymore, and nothing really goes out anymore. Um, so building such automations via very simple tools like Zapier, NN, etc., is easily possible in a data protection-compliant way, no problem. Um, if you work on your own server, you can also host ChatGPT there. Um, then there will be no data protection problems there either, etc. That is, uh, that is all possible. So, everything I am showing you now is possible in a data protection-compliant way. Um, that is possible. Don't worry about that. Yes, and this step two is now, um, really about linking tools like, for example, Langdog with, uh, other tools and then doing something with them. Um, and now I will switch the browser. That means I will go out of this, out of the screen sharing, and go into here. Yes, you should now see Langdog on my screen. Um, here it says "Site," that's a logo that I, oops, that's a logo that I added myself. Site AI Workspace. When you log into LDOG, it says LDOG. Yes, then you are working with LDOG. What is LDOG actually, and why are we doing this now? Why are we using it? I actually wanted to get to the start page. Shame. Uh, so. Langdog. Um, Langdog enables you to use ChatGPT, Google Gemini, Claude, etc., in a data protection-compliant way. Why? They have bought an enterprise model of ChatGPT. Uh, if you take, uh, or want to buy and use this enterprise version, it costs several thousand euros per year. Uh, then you have the option to tell it, or tell ChatGPT OpenAI, please, um, process my information, and what Trebt does, uh, please do it on European servers. And if it's on European servers, then it is, um, in this sense, also, OpenAI had other ISOs, etc., then it is all ISOs, then it is data protection-compliant. Yes, and because these individual accesses to the enterprise version of ChatGPT are super expensive, um, Langdog has simply bought this access, also for Google Gemini, also for Claude, also for other AI. Um, it uses this, uh, processing on European servers, and offers you the option to simply register for, uh, €20, yes, €25 per person, per user, and offers you the option to access it and use it. And that is completely data protection-compliant. There is not only Langdog, there is also My GPT and others, uh, that do this, uh, that do it this way. Uh, so they have actually solved this data protection problem for you. And I am using it now, I wanted to tell you that, um, I am not using it now because it has a lot of data protection possibilities, but because it offers possibilities to build such assistants. This is essentially exactly what we built earlier with, uh, with ChatGPT, uh, as a custom GPT. Um, you can build these assistants, but you can also provide them with integrations to your tools at the same time. Here there is a range of tools already displayed, other tools that you use, on Office, etc., uh, you can add yourself. I have done this with Fireflies, for example. It's also relatively easy. The AI tells you, uh, what you need to do, and then it works. I have linked my Gmail account here, I have linked Google Calendar here, uh, Google Drive, etc. You can also link Outlook here and, and, um, so, and this is always, it might look complex or something. Uh, I can't show it because everything is already set up for me, but you just click on connect. Then a page opens, you have to log in once, and when you log in there, you are connected here. Then you also tell it, yes, you have these and these and these rights to work with LDOG, and then you are connected. But it is not a security, uh, problem at first. Um, of course, a company might need to coordinate this with, uh, data protection, or also with the IT department, but fundamentally, there are no, no problems, and your login details, etc., are not stored and are not passed on. It is simply a connection via these APIs with each other, so that you can use both tools together. Yes, and then there are nice things like so-called MCPs. Um, all these, uh, all these, yes, tools behind it, uh, are controlled via so-called MCPs, because an API, an interface, actually works by you, uh, saying via code, in principle, I want from HubSpot, for example. Uh, I want you to search for this and this name in HubSpot and then give me the email address of the person. Yes, that would actually be, like, three queries. These MCPs, however, are the connection between an API and an AI. I hope it's not getting too technical, or there are not too many new terms. Um, the MCP, the AI understands the API with it and understands you too. That means you simply say something that you want to do via this interface in the other tool, find out what data you want, what you want to adjust there. You write that, you say that with text, the MCP understands what it has to do in the other tool and gives it back to you. That means I have, uh, I have brought an, um, an example with me. So. And I will enter this example here. I can now, for example, um, we use HubSpot at our company, and I can use it via this. So I can now say, create a new contact in HubSpot, Alexandra Webinarmann, with this phone number, um, and this email address, and add a note that I spoke with her today and that she is looking forward to the next practical webinar. Also enter the first and last name correctly, and assign Nieders Möllenkamp as the responsible, uh, employee. And now I just say HubSpot, because HubSpot is linked for us, and now say go. And what the AI is doing now, you can see it step by step, is, it has a follow-up question. Um, now it needs my owner ID. I didn't expect this follow-up question. Search, uh, find the owner ID for it yourself. Yes, it finds that itself in HubSpot. And, um, so, and now the question is, is it allowed to search for the HubSpot owners now? Yes, it is allowed. You can set it up so that you have to confirm it every time. To run it truly automated, you would no longer do this confirmation. And now, uh, it says, should I please create this contact with this email address, with first name, last name, with the phone number, um, etc. And the owner, that's apparently my contact ID at HubSpot. Um, should I create it like this? Yes, please do. Then, uh, it says, we have a lead status that doesn't exist and the phone number is invalid. My suggestion, I will create the contact now without a lead status, without a lead source. Yes, do that. Yes, once you have set it up, oh God, once you have set it up so that it works, it will no longer ask you all these questions, and you won't have to confirm it anymore. So, I could have turned off the confirmation now. I didn't do that to show you what happens. Right now, uh, it says, I'm also adding a note, and now it has added the note, and now I can also say, for example, give me the link to the contact you created, and then we'll see how it actually looks in HubSpot. This is the link. Let's go there and see, Alexandra Webinarmann has been created. I spoke with her today. She is looking forward to the next practical webinar, and this is her phone number. Yes. Um, and that's how these MCPs work. And now we'll go a bit more into the actual, uh, yes, real estate exposé, which we have here. And, um, we'll do the following. We'll create an assistant now, because these assistants can also access these MCPs, these other integrations. Create assistant, create new. This actually always looks quite similar to what you know from ChatGPT. Um, and now I'll take the pre-made prompt, which, of course, was not written by me, but by my friend ChatGPT. Um, what exactly? Something is missing for me down here, which is surprising me right now. Yes, live and practical webinar always means there can be topics. Okay, it took the top section. Sorry, I have to copy this from bottom to top in a completely stupid way for me for it to work. Sorry, that, uh, so, I have actually figured out how to do it. Okay. Um, but I have a solution for it. Don't worry, we're ready, so ready. I'll copy this out. And this must work now. I don't understand why I can't copy this prompt. Um, this is very embarrassing for me right now. Uh, it shouldn't be like this, but, uh, luckily I still have the prompt, of course, with ChatGPT, which I just need to search for briefly. Um, yes, sorry, and honestly, I'm very bad at doing two things at once. It doesn't output it for me. We need this prompt, what the prompt does. Okay, now I have it. I just need to copy it from something else, apparently. Ah, that doesn't work. Okay. Um, then we'll use the, um, I'm very sorry. I would like to show you how to build the whole thing now, but of course I already built it yesterday, so that we can at least use it now. Let's go into assistants and see. Uh, it should be called something like exposé. Um, Exposé Scanner, yes, this should be it. Edit Assistant. It looks exactly like what we just saw, only with the entered prompt, which I also copied in there yesterday. So it still worked yesterday, but okay. What does this tool do now? Um, we'll do it like this, it has access to my Gmail account, and, um, if I say start or anything else, it checks the last emails I received, for example, with the title "Exposé." Yes, if I received an exposé as an email, then that would be the email it uses. And then it checks, uh, to extract the important exposé data from the exposé attached to the email, and then it checks or compares this with the acquisition profile. The acquisition profile is also stored in this prompt here. This could also be done differently. Then it evaluates whether the property generally fits, and the email, and then it sends an email, uh, to, ideally, my email address, and then says whether the property, or what has just come in, matches what we need or not. That means these steps that we did with ChatGPT earlier, and to see, it's again very, very extensive. The third step, which we did with ChatGPT earlier, it now does automatically. Yes, for this, I have added actions, Gmail is linked. Yes, it should always ask for confirmation before sending an email and writing, of course. Um, and that's actually all we need for now. I always recommend selecting the model here. I said earlier that via Langdog you have the option to use ChatGPT, Google Gemini, and other, uh, other, yes, AI tools. I would recommend Claude with Sonnet in the reasoning variant, because an exposé can be somewhat extensive, and for analyzing something like this, um, then, uh, Claude is already the best fit. Yes. And now I say start. Simple, very simple automation. The user has written start. According to my instructions, I now have to search in the email inbox. So, and it's doing that now. Now we see here, it's searching for emails, for the last email with "Exposé" in the subject. It has found several emails, the last one yesterday evening. Um, and then sees this attachment here. So that's the exposé. And now it says, I found it. Cool, now it has read it. And now it extracts the data for us here. Yes, all the data that was entered in the exposé. In principle, like before. Price not specified, features fully developed. Now it compares it with our acquisition profile, which in this case, as I said, is stored in the prompt. Says overall suitability fits. Yes, is that what we're looking for. What speaks against it. We don't have a purchase price right now, that's missing. Uh, conservation figures are only forecasts. Um, and there's a large range in the figures, actually. Yes, so nice. And then it creates this email here, um, to myself, and says: "Listen, uh, we have a pulse rating for this exposé. Send it to me." Um, it's currently creating the email, we'll see that in a moment, and then we can say, okay, send it. Um, so, here it's written in HTML, uh, and then says here, property size, what fits the profile, what doesn't fit or is unclear. Uh, and a recommendation for it. Yes, um, alternatively, confirm this, and now it sends me this email. Alternatively, you could also directly say here, reply to the agent, we still need this and that. Um, or, uh, or add the whole thing to a list, in a table. Um, all of that is possible with this. It works. However, if you want to extract the data into a list, it's a bit like, if you have a hammer, then everything is a nail, whether it's a screw or something else. That's a bit like the AI point. You simply use the AI because you, uh, because you have it available. Um, if you, if you have extracted the data with AI, you would then actually, uh, enter it, for example, into an Excel table or a list. For that, you would use other tools like Zapier or Nend, which we will now, uh, go to or enter shortly. Um, but this way you have sent this email, and you could also build this with workflows in Langdog. This goes a bit into this, or not a bit, it goes strongly in the direction of Zapier, NN, so that you say, every time I receive an email, first of all, every time I receive an email, check if it's an exposé, if there's an exposé in it. Um, and if so, then extract the data, put it in a list. Um, then, uh, you would write me an email and make a short summary. Is this suitable or not? Um, yes, and that's it. So, you could also build that here with a workflow, so that it's always triggered and happens automatically when you receive this email. That would also be possible with this. Yes, but we have taken the example, because I believe it is presented a bit more vividly, and, uh, you can see it a bit better. Um, exactly, so. And now I have to, uh, I We see, we have already been in the other browser. Why? Uh, I'll show you that shortly, or I'll show you that shortly. Um, we will now go into, uh, the presentation again, because we will now move on to step 3, and step 3 is, uh, to build this exposé process as a workflow. And we'll do that in this case with Zapier, and I've done that once before and will show you what the whole thing looks like and what Zapier looks like. I use Zapier for that. I've familiarized myself with it. I would recommend something like Zapier or Make.com for that. Um, and this process, which we did with the AI earlier, where everything was a nail, yes, because we had the hammer, um, if you want to build it meaningfully, you would actually build it like this, and here you can already see what happens step by step. Uh, and here we can set and build the individual steps one after another. Uh, if I click on "Edit Draft" now, you'll see, uh, how you do it here. So, this is the trigger, for example. Um, you can also very easily with a few clicks, uh, link ChatGPT, yes, ChatGPT too, but also your email inbox, um, etc. And it will be triggered, so it works automatically when a new email arrives, and in the specific example, in this new email, uh, in the subject, the word "Exposé" is present. Just as an example. Of course, it can also be done differently. Clear, or should actually be different, because not every exposé has "Exposé" in the email subject. And then you can test it once and see what, uh, what the, uh, what the email is that it receives now. And now you see here, it has received a snippet of the email. Uh, and this here is, in principle, the email that it sees. Yes, with many different things, things that I don't necessarily understand. But somewhere we still have the email text, if we don't find it quickly, um, believe me, yes, you have the email text here, you have the recipient, the sender, and also the attachment, the exposé as an attachment, all included here. And then the whole thing actually works like this, uh, that you simply click on "Add Step" here and then say, which tool should work next, please? If I were to say, I want, um, I want, uh, to upload this, this is what we have already done here. I want to upload this to my, uh, Google Drive. Then you click on it and simply say Drive. Um, so to event, right? Upload File, so continue. And then you select here, so, for this to work quickly here, or for this to work quickly, you won't learn it 100% now, but these are, of course, things, look at it, um, and see how it works for you, and then it's relatively quick. We will now select the Google Drive, the folder in there, and now I say, Test in Webinar, let this file be named, and it should please be converted from the, uh, Doc into a Google Document again. Um, oh, no, sorry, I did it wrong. Under File, you would now use a dash to simply refer to the attachment, uh, from the first email here, we can refer to it, then this here, you would enter that here. And the filename is then Test from Webinar. So, continue. And I can also test this here. And, um, I'll show you, um, I'll show you the folder as it looks right now. Yes, it looks like this. Yes, Expos Hilden and Selfia Table are currently in there. If I test this now, Test Step, it's working. And here we will soon see another, another file. It's still working. It's finished now, it has uploaded this file, and now we have the file Test from Webinar with our exposé. Yes, it has now uploaded it automatically, and you set it up so that it always happens automatically when this file is uploaded. When, when it finds a new email that matches these criteria. And then this works automatically, completely automatically. I'll delete it again, because we already have it here. Um, then you would pull the contents of the document out here again, because of course we need the text. And as the next step, um, an AI, an AI is already integrated here. In this case, OpenAI is working with ChatGPT 5. By the way, you don't necessarily need your own ChatGPT account for this, because it's already integrated here. So you can use it automatically in this GBT5 variant, and then you send the content of the exposé, which you get via this, and a prompt to ChatGPT. You are a highly precise real estate analyst, and your only job is the clean data extraction from this email. This is the subject. Everything comes in automatically. Uh, sender and here the, uh, and here the, uh, exposé, and then there's a prompt in here again, which you can imagine, which I had ChatGPT create for me again, which simply says, please extract everything, and it does that now. We can also test it once, so that you get all the information structured like this and get it out structured like this, as we see it here now, because I have already, uh, already tested it yesterday. Yes, and then the next step would be to say, listen, yes, okay. Um, and as soon as you have extracted this data, I want it to be entered into a table. We also have this table here, yes, with these column names, and I want it to automatically enter everything here. It hasn't done that yet. This is also from yesterday. Yes, but, uh, simply link Google Drive again, in this case, link the table, link the worksheet, um, and then assign what it has extracted to the respective areas. That's how you do it, it's safer than letting Langdog do it automatically as before, because we can specify exactly this assignment. Yes, street and house number it doesn't have, uh, can't specify anything. For others here, I haven't specified anything. Property size, always assign that directly, then continue, and I can also test this here, and if I test this now, then you should see it here, yes, it has added it as another column, and this is how you would always, when an email comes in, so again, every time an email comes in that contains the word "Exposé," the exposé will first be uploaded to a folder. Then we get the contents out. The contents are automatically extracted with an AI, and then stored in a table. And these steps, of course, can be as large as you want. You can also, because we had the question afterwards, include a research tool and say, yes, we don't have a development plan right now. Please research, or we don't know if there's a development plan. Please research a development plan and also enter that into the table. Then you could set up a scorecard here, so that you say, okay, um, now you should also, uh, please, uh, please check, does this fit, how strongly does it fit what we're looking for? And then you set up a scorecard here, and then the AI says, "Yes, it fits 30%," then I'll give it 30 points or something. And in the end, you'll have a precise evaluation in this table of how good the exposé is, uh, tailored to your acquisition profile, with, so not 600 out of 1000 points, and you say, m, that might not be quite it, but with this you can build this entire process for yourself. Yes, and, um, now we come to the last tool, step 4, with a bit of luck with time. I've just cost myself a bit more time than I wanted by not being able to copy the PR. Um, the last tool for that, that's then, uh, that's then N8N. It looks a bit different here than Zapier right now, and it also looks very technical. So, it's something where, as I said, I'm digitally interested, but somehow this is something where I run away from at first glance, honestly. Yes, but you can also build these workflows very easily, and once you've understood it a bit, it works even easier than, uh, than Zapier, for example. And why do I have all this in another browser? Well, um, it is quite possible for you, uh, that you, um, yes, then with ChatGPT, oh, I'm not sharing the right screen at all. I'm very sorry. So. Yes. Uh, now you see the correct screen. Um, and it is now very easy to get involved in how this whole thing works, simply via ChatGPT. That's why I have this browser open. I can, uh, so it's called the ChatGPT Atlas Browser, and I can simply ask ChatGPT on the right side of this browser, um, and work with ChatGPT. And if I enter a prompt here, like this one: "Research today's date and find all information to give me a very precise guide on how to get an assessment of the condition of a property based on multiple image attachments in an email with N8N, via ChatGPT, and have it automatically sent to me via email." So, a slightly different use case that I've taken. We receive images via email of the interior of a property, and then we want ChatGPT to see these images so that ChatGPT can then send me an email via this and say, yes, um, maybe I wrote the email to myself at the beginning with the images. Uh, I wrote this email to myself and am now receiving a reply directly with the, uh, with the evaluation from ChatGPT, having everything in one tool, maybe on my phone, and having this evaluation directly. I can build this very easily with this, and ChatGPT helps me with this. Theoretically, there's also the N8N AI at Zapier. Um, at Zapier, there's Copilot. That means I would simply enter this prompt into this AI. Why we're not doing this today is because, firstly, this AI assistant is not always correct. It builds all of this automatically for me, but with many hiccups, with many problems, and if you don't, uh, know how N8N works, and if you haven't built it yourself, then it's of no use to you, because then you can less, uh, assemble it yourself for yourself. And that's why I would always recommend not using it this way. The second point is, it's relatively expensive to, uh, build this process like this. At Zapier, it's a bit cheaper, but it just doesn't work really well yet, I have to say. Yes, that's why I would do it this way, with the ChatGPT Atlas Browser, and you have the option to have it output directly what you actually need, and it will guide you very, very precisely through what you actually need to do here. Gmail trigger, listens for new emails. Yes, that's this here. It looks like this. You see, it's a bit, it looks a bit different. It looks very technical. What it's doing here is, in principle, after you've linked your Gmail account again, which is easy to do. Um, it looks at the attachments, so the images that come in. But what exactly you need to enter here, yes, ChatGPT always tells you. And the browser is good for this reason, because if I ask a question about it now, ChatGPT sees everything that is visible in the browser. That means you can immediately say, this doesn't work, why doesn't it work? Where did I enter something wrong? Research that. And then it will tell you, yes, you need to change this and that, you need to adjust this and that. And then, um, it works, um, very, very well. One or two times it will tell you something you need to adjust, which then also doesn't work. Then you say, that doesn't work either, look at what it looks like. Then it finds the correct solution, and with that, you will be guided step by step, really. So it tells you step 1: Set up Gmail trigger, that's what we have here. Um, set filters, recommended with a specific label, and so on, you build it up step by step. Then here, the second step is a script, a JavaScript, where, um, the images are probably separated from each other. What exactly happens here, I have no idea. I can't say 100% because this prompt for it and where I need to enter it, the AI tells me. Yes, if I were to sit down now and look at it, I would probably understand it, of course. Uh, or I can also have it told to me, right? Tell me how it works, and, uh, and, uh, would understand it. Yes. Um, but for the first test and for the first time working with it, to work with it, um, yes, this is splitting image attachments into individual items. That's exactly what we see here. It says, insert JavaScript code. You take this code, copy it in here. Execute Step. You can always have the individual steps shown to you how they work. It's behind it now, you can't see it properly, but apparently it worked. And then the next step is with ChatGPT. with a prompt that is in here. I'll scroll down a bit here, ChatGPT itself also gave me this prompt. I enter the prompt here, um, linked with my OpenAI account, and now it sees the individual images. Execute Step. Looks at them and gives me a description of what the object looks like from the inside. Here it comes shortly. It takes a few seconds until it has really done that for every image. But we are essentially telling ChatGPT six, eight times in a row, depending on how many images you have. Um, here, you are a building surveyor. Uh, you are a building surveyor. Please look closely at the condition and give me a condition description. We do this six times in a row. It does this completely automatically until it's finished at some point. And then GPT says here, all image analyses, merge image analyses, again with such a code and with this GatbT model, so first all texts one below the other, and then it should summarize it again. Yes, and then the whole thing will be finished. Do we see it yet? No, here it looks like this, also relatively technical for me, or at least from my perspective. Um, it looks like this, it's passed on here and will be sent to you as an email in the end. As I said, the best, uh, the best help is to do this with ChatGPT on the right side. The hint came, the browser is only available for Mac. Yes, that's correct. Um, but otherwise, use, go, yes, put N8N on the left side of the screen and ChatGPT on the right side. And what we're doing here is essentially just, it always gets a screenshot of what we see here. You can, uh, you can also very easily, uh, open ChatGPT, sorry, ChatGPT on a second, uh, screen next to it and take a quick screenshot, drag it over, say, this and that doesn't work, how do I do this and that, and it will help you further, and then this is your best partner to really build your first automation like this, and even how do I start now, how do I get started, where do I even need to register, and things like that, just ask ChatGPT, it will help you, it can do that, and if it doesn't work the first time, then the second or third time, if you, uh, describe the error more precisely in the end. Yes. Um, that's how the work is done, and I hope this was a, uh, for you, for most of you, a very good introduction. So, I hope of course that it was a good introduction for all, uh, of you into the question of how something like this works. Uh, how can you actually build and do something like this? And, um, as I said, it's an introduction. Um, you will, um, get more in-depth information on such things in the coming weeks and months. Um, feel free to check the Site website for more webinars that we, um, that we do in that, in that, in that context. Uh, feel free to check Site on LinkedIn. Uh, there too, we regularly show what we do or what we're doing, uh, what the next webinars are, and of course, a lot about Site. Uh, my contact details, if you have questions, now afterwards, um, please do. Then scan this code, enter your email address. You will then receive my contact details, will not be put on any lists or anything, don't worry. Um, you will simply receive my contact details, and, uh, you can then gladly write me an email or, uh, write to me on LinkedIn or something, if you have questions about it. Um, this has already gone a bit more into depth, yes, than the other webinars that we've had, that we've otherwise had. Um, I hope you were able to take something away as a first introduction to this. Otherwise, as I said, feel free to ask me, we will do more in this area and offer and show more. Um, and if you also have ideas for future workshops and webinars, we would be very grateful, um, also for how we shape it. Uh, we are very grateful for that. Yes, okay. Then I will conclude at this point with a slight delay. I really need to look again why I couldn't copy that. That annoys me a bit, uh, with a slight delay. Uh, that was at this point. Uh, I was very happy that you were all here and, uh, so a very high number of participants here. I'm very pleased, and, uh, yes, until next time. Thank you.