Transcription
So, automated acquisition and property due diligence thanks to AI is the topic today. It is, uh, like the last seminars, webinars that we've done in this format, a practical webinar. So, um, you can already see that I have one or two more tabs open. Um, this will not decrease as we, uh, look at what we're actually doing here. Um, so we'll go into very different tools in a moment and see how these tools can help you in acquisition, uh, and property due diligence in the end. And, um, yes, as I said, my name is not, uh, as it says here in the picture right now, Matthias Züge, but Nilas Möllenkamp. Um, I've been part of the real estate industry for 7 years, a long time at BNP Paribus as a project manager for digital strategy. I'm not a developer, not a techie, not a coder. Um, I can't do any of that. Uh, honestly, I have no idea about that. When I see code, I run away. But, um, through all the AI, I've come to the point where I can build programs myself, build automation myself, put things together. And, uh, I've only really been doing that for two years now. Um, so some time after I started at Side, through whom we are holding this, uh, webinar today, or rather, not about Side in terms of content, but, um, yes, Side is the, uh, company that invited you to this webinar. That's how you have to put it. And, um, I've been able to teach myself a lot through AI about how to automate processes, how to build processes with, um, with AI. And that's what I, uh, want to share with you and give you today, um, and show you what you can actually do today when it comes to property and acquisition due diligence. And in my work, wanting to bring this into the real estate industry, I am, among other things, a key speaker and a guest lecturer at ADI, CIA, BDB, IRWAPS, um, and yes, I also provide AI digitalization consulting, uh, in one form or another, simply because, um, because I was able to quickly teach myself. Side, by the way, is the German AI startup of the year and actually brought the topic of AI, um, to my attention. So, and what you see here, by the way, is all, or there is one image in this entire presentation, which is this one. It's AI-generated, by the way. Perhaps you can see that it's relatively similar to the image, uh, that you saw when you registered for this webinar. Uh, AI has indeed progressed to the point where I can give an input prompt and upload this image of myself, and a really good image comes out of it. Um, I've included that here once. What the image is supposed to convey, however, will probably only become clear with what you, uh, what you will then also read. Um, artificial intelligence, concrete benefit, real examples. Uh, when we do these webinars and, uh, when we give seminars, hold them, um, etc., it's always about showing you very concretely what you can do today, what you can use today, um, because we want to, um, encourage you to say, yes, uh, with AI, I can really automate a lot of my daily work, of what I do every day. That's what I do. That's what I contribute a bit at Side, and, uh, it's really not difficult. That means we show you, or I show you in this case, what you can do, how you can do it, how you can implement it, what the corresponding tools are for it. Of course, it's still the case that you'll probably have to sit down for another hour to recreate things, to replicate them, uh, to adapt them a bit to yourself, but then you'll know directly, okay, I can implement it like this now, and, um, I can use it like this now. And I promise you that there are things here today, um, that you might need, uh, an hour to adapt, to make them usable for you. And afterwards, um, you can transfer it one-to-one, and it will save you 30 to 40 minutes per week, definitely, and immediately. And it's very simple, and that's what fascinates me about it. Little input in time, a lot of output. And therefore, because we've unfortunately lost 4 minutes at the beginning here, um, we'll now go straight in, and we'll start, um, above all, with an AI tool that you, or that most of you know, which is ultimately ChatGPT, because, um, in acquisition due diligence, in property due diligence, it's very, very easy to, um, give data to an AI, to a large language model, i.e., a text-based, a text-based AI like ChatGPT, and, um, with this information, then, uh, extract data that you really need. What is AI particularly good at? You upload data, a lot of data, a lot at once, and you get exactly the data you need very quickly because you entered it that way beforehand. So, extracting data from a exposé is, with AI, actually problem-free, problem-free. And that's the first step, which I'll show you once. Um, I've had various exposés created. I'll show you one of them. Um, these are now fictional exposés, so Waschauerstraße 45 in Berlin, uh, is not currently, um, a building plot. Yes, uh, I've at least checked that again beforehand. If it's currently being marketed as a building plot, then something's not quite right, I think. M and you see here in this exposé, uh, a number of details that could appear in a classic exposé. I've simply saved the whole thing as a stock, um, and to extract it like this, because an exposé can easily be, uh, 10 pages, 15 pages, depending on the situation. Um, you'd first have to read for 30 minutes, maybe 15 minutes, maybe you'd check one or two things, uh, so that it takes an hour, um, to really get the relevant data for yourself, uh, to get it out. Excuse me. And, um, the whole thing is much easier with AI, and we use, we use ChatGPT for that, and, um, for that, I'll now enter the prompt that I've prepared here. Um, you are an assistant for real estate acquisition due diligence. I will upload three real estate exposés to you. Uh, please read all documents carefully and extract all acquisition-relevant data into a clear table. These are then, for example, these here. Um, the prompt for that, what I'm entering there now, you can easily generate it yourself from ChatGPT or from another tool. So, I don't sit here as it looks now and ensure that, uh, everything is really, um, perfectly written, that, uh, there's a perfect overview, um, overview of the data to be extracted, but, uh, I let the AI output something like this directly itself. And now, uh, I have this prompt, upload the three exposés. I showed you the one earlier, and, um, practically click on go, start. And now a longer thinking process begins, because we've uploaded three, um, files, it takes the AI a bit longer, uh, to, um, to review them, or rather, to, um, verify the information within them to some extent and ensure, uh, that the tool has understood them accordingly. And, um, yes, it still needs a very short citation, of course. We had that, uh, in this, uh, in the prompt, that it should at least say where, uh, it got the information from. And we now say, for example, fourthly, summarize each of the plots in three to five bullet points and, uh, yes, sometimes you have to wait a bit at this point, sometimes it's faster. I could also try to make the whole thing faster now. I would open a new chat for that and click on Instant here at the top of ChatGPT, then the data, or what comes out, is not 100% as good as before. Um, but it would be faster, which, honestly, in our case, um, is sensible. And then it's already finished and extracts exactly the data from each individual exposé that I need. Address, plot size, GZ, GZ, buildability, and so on. Yes. Um, the example we have now is simply that we are looking for building plots in the end. That's why it's mainly about the plots here. Uh, price per, uh, square meter was calculated, excuse me, calculated, um, and already calculated. And, uh, that's the very simple AI extraction of this information with the corresponding short summary, uh, bullet points that we have here. Yes, and that's the first step, and it's very, very simple. And the second step, uh, is as follows: we do the whole thing via a so-called Custom GPT. Uh, what is a Custom GPT? A Custom GPT is there so that you don't have to provide certain information anymore. And by certain information, I mean, for example, the prompt here. That means we, um, can build the AI to a certain extent so that it does exactly what we, uh, what we actually, what we've given it once, what we actually absolutely want from it. So, with ChatGPT, other large language models also have these capabilities, but with ChatGPT, I click on GPTs here, click on create here at the top, um, enter a name here, exposé, uh, evaluator, something like that. Um, and then it's important that I enter exactly what, uh, the AI needs to really understand what it should do here in instructions. And in this case, uh, we'll do it a bit differently. I'll make it a bit bigger. Uh, in this case, I've of course had the prompt generated by ChatGPT again. The prompts for that are always better when you have them generated by ChatGPT than when you write them yourself, because ChatGPT knows exactly what ChatGPT needs. And if I tell ChatGPT with two sentences, output the prompt for this, I want to do this, and then this comes out, um, then we'll definitely get somewhere with that. Yes, the prompt, uh, is, uh, a specialized assistant for real estate acquisition due diligence. Then it contains the fixed acquisition profile of Möllenkamp GmbH. So, I've given myself, uh, my own project development company, and have also stored my acquisition profile here to some extent. Object criteria, usage, storage criteria, uh, investment and return, process, etc. And then come your tasks with this GPT: extract data from exposés, i.e., analyze uploaded exposés, PDF or text, and extract all relevant information. We had that before too. Um, comparison with the acquisition profile, identification of open points, structured presentation of results, and additional rules. Um, yes, come in as well. And now I click on close and say, create only for me. So, uh, that we don't make the GPT publicly available. And, uh, now it saves the whole thing, and what I then only have to do is display the GPT, and, uh, what you see here, uh, first of all, the GPT was automatically filed here on the left side. Um, here at the top it says Exposé Evaluator, that's the GPT we created. And here you see the possibility that I now have, that I can upload these three exposés here without entering a prompt. Um, my acquisition profile is already stored, and now it says it has evaluated the three exposés, this is again in this instant mode. So, if you let the whole thing think a bit longer in this thinking mode, the results are a bit better, it must be said. Um, but no, we're doing the faster version now. Uh, now it has extracted these property data here, all in a clear table, as we would like it or need it. And now it compares it with the acquisition profile. Exposé 1. Um, we have certain points that match. Um, evaluation here partially suitable. Um, Exposé 2 partially suitable, tendency red, because it's a listed building. We didn't want that in our acquisition profile, for example. Um, Exposé 3 in Spandau, not suitable, red, uh, because no residential building use, sorry, is permitted. Um, and that, uh, and that we have then, for example, in the sample acquisition profile, uh, stated very clearly like that, that we want that. And, um, now come open points and recommendations. We would still need a land register extract, contaminated sites, expert reports, uh, and development plan study. Um, recommendation yellow, despite prime location, too small for profile, but could be interesting as an exception. Um, these are then things that, uh, are then also output here accordingly. So. Um, now we have created this Custom GPT. Uh, now I'd like to show you one more thing, um, which also gives a brief look at, uh, at Side. Um, what you get with Side, uh, is at the push of a button for every plot, um, the development potential that we would have, right? So, I've already said, it's not a vacant plot. It's now a completely arbitrary example. Um, but this plot has a development potential of almost 4000 square meters. Side. And, uh, you can also, um, yes, download so-called reports here. I've already done that for this plot. Can I access it like that? Yes, uh, I've already done that for this plot. And, uh, you see this report, for example. And this report gives you, um, because it knows the development potential, when it comes to vacant plots or also developed plots. The report gives you much more information, uh, about what, um, the real estate agent tells you, the real estate agent tells you, right? We currently have a building on it, this building is from these construction years, the building is this big, and and if you now take such a report, um, then you can, of course, also very easily, uh, incorporate it yourself and, uh, into this, into this property due diligence, by now, for example, building a GPT. So, we'll do that again now and say, um, exposé, uh, buffer plus Side. Yes. And, uh, now we'll build in a corresponding prompt as instructions again, na so. Um, what has essentially changed is that it now simply says that, uh, the exposé, uh, also, uh, a Side report is added. And, uh, now we create the whole thing, and now we upload this one exposé from Warschauerstraße in Berlin. So, we upload that, and then we also upload the Side report again, and simply click on Let's go. Um, the point is, of course, it also works differently, so, um, or rather, this functionality behind it. You don't necessarily need a Side report for that. You can also use a Price Report or any other tools that you normally use. Primarily, I want to show here, you simply upload an exposé, uh, and further information about it, and if you've previously told the GPT that you'll upload further information, then this GPT can also clearly distinguish here. In the exposé, the plot size is 1200 sqm, in the Side report 771 sqm. At the same time, the GRZ and GFZ are, uh, it's said that a GFZ of 3.0 is possible in the exposé. However, in the existing building, we already have four, and, uh, the GPT also recognizes the Side report from, uh, uh, yes, from Side immediately and sees that we actually have a potential of up to 6.09 for the GFZ. Um, and that's also indicated here accordingly. That means, if you now have an exposé, if there are plots, if you have the exposé and a report, uh, and create a Custom GPT for it, then you get, you see it here, very clearly, um, all the information you might need for this, uh, plot directly, um, based on this, a very simple purchase decision, or at least a decision whether the plot, um, goes to the next step for you or not. Um, so, expand GPTs, in the best case, of course, with a report, but expand GPTs, uh, and say, here comes additional information, and then calculate with it. Yes, that's relatively simple. Um, exactly. So. Now we have, uh, further topics that I want to discuss with you here. Um, I just need to check briefly, uh, where we were, because these prompts that I write down beforehand are always so long that I have to scroll a bit. Um, next point. Uh, I'll first show you individual steps, um, which you can then also easily combine in the end. That's the last point where we'll go next. These individual steps also mean that, uh, we've now worked completely with text and, uh, a bit of data and numbers basis, and on this basis, we've made extractions. Uh, but, uh, AIs are also very, very good at recognizing images and, for example, providing condition descriptions. And, uh, what I've brought with me is this example. Um, I've uploaded a picture from the inside here, which I found somewhere on the internet. Uh, and I've basically told Google Gemini, another large language model, uh, that I want, um, you to describe the possible condition and the condition as well as possible damage, and what comes out here, so the tool is extremely good at recognizing photos and also recognizing things from photos. That means we see here, uh, yes, it looks like it needs renovation. Well, I would have seen that too. Um, we have water damage, moisture on the ceiling. Let's look. Yes, up here it's going in that direction. Um, we have mold formation, we have worn, damaged plaster on the walls and ceilings, worn wooden floorboards. Basically, it says here that everything in this room is bad that can be bad. Yes. Um, but what I actually want to show you is that an AI can recognize this. So, if, um, you receive a property where the real estate agent makes it very clear, uh, how beautiful everything is, but then, um, but then in case of doubt, you say, uh, yes, I'd like to have that checked again, but you don't want to look at the photos yourself, whether on ImmoScout or elsewhere, or in the exposé. Um, then it's very, very easy to have the whole thing done by the, uh, AI as well. And, uh, here too, I've prepared something, um. I won't let this run automatically now. It probably took a minute at most. I found a few photos of a property from, uh, from ImmoScout, somewhere in Berlin, that I found by chance, and so a few photos from the inside, right? I think the quality, as it's displayed to you now, is not as good as, uh, as I uploaded it here, it must be said. Um, but you can see a bit of the building here, and then with a simple prompt, um, you are an experienced building surveyor, analyze the provided images of the apartment building, uh, assess the visible building substance, and structure your answer clearly. Uh, then something comes out here. And that's not bad. Recognized defects, recognized or possible defects. Stairwells show significant signs of wear. The handrails are heavily scratched. Um, windows appear to be old. Uh, general condition. That means, um, if you have an exposé, there are pictures in it, then upload the exposé or the pictures individually to Google Gemini. I would recommend that for that. ChatGPT can also do that quite well. That's okay. Google Gemini is a bit better, right? So, gemini.google.com, google.com is the corresponding address, and I believe that it's also still available free of charge to a certain extent right now, but then you upload these images, this data, um, think of a good prompt or have it thought up by another tool, and say, yes, how should I actually assess the condition of the building now? What would also theoretically be possible is to say, how do you estimate the costs that arise based on the fact that we are, for example, renovating it? And it would also output something. Um, if you enter the exact same prompt with the exact same images seven times, you'll get a slightly different result seven times. That's just how it is right now, unfortunately, but it always goes in the right direction, or it always goes in one direction, and with that you can, um, work very well initially to do this first very quick preliminary check or this first very quick check. That means, um, also remember with AI that it can analyze images for you. Even if you have 20 images, you can upload them all at once and, uh, yes, see, uh, in which direction the AI evaluates the entire property accordingly, um, yes. So. Um, then, oh, yes, this was the property, sorry, uh, I just found a link where I wasn't entirely sure where it came from. But then let's go to the next step. What we've just done was, uh, extracting data, uh, and checking, uh, this existing data against the acquisition profile. Um, and of course, also assessing the, uh, the data we have, in principle. Yes, and, um, now we go one step further, and now, uh, it's about using artificial intelligence to research further data and further content. Um, that means expanding the available data for due diligence using artificial intelligence with the tool of deep research. Regarding the deep research tool, um, if you've already attended one or another webinar and lecture here, um, you'll know it. I always like to focus on it because it has changed my perception of AI a lot. But ChatGPT, for example, has a function, it's Deep Research, which translates to deep research. Gemini Google Gemini also has this function. Perplexity is also a large language model. It has, uh, the whole thing is called research. Uh, it simply goes into more detail and researches information for you. What information does it research? In principle, everything you can find on the internet. My first start with ChatGPT Deep Research was that in Düsseldorf, Oststraße 91 and 93, uh, they've been vacant since I've lived in Düsseldorf, for 7 or 8 years, and honestly, it annoys me a bit because it's very central. Um, so there are two plots next to each other, they are very central, you could actually consider a zoning closure there. And, uh, that was the first thing I ever built into ChatGPT Deep Research. Um, because I simply wanted to know what's going on, why nothing's happening there. And ChatGPT Deep Research took 20 minutes and gave me everything it could find and know about it on the internet. And, uh, I know now that the first people lived at this address in 1871. I didn't even ask for that. Um, but Deep Research found this information and thought, okay, then I'll just share it. Yes. Um, in an overview where I could have also said myself, okay, I'm not reading any further here, this is not relevant for me right now, but, uh, no, this information was extracted and found somewhere from some websites. Likewise, um, I was told or it could be told to me that there used to be a single-story building there. This was found from various sources. Uh, but above all, Deep Research also found a, uh, presumably free advertisement that was somehow put in the mailbox from 1970 or so. Um, where, in turn, there was advertising in it, i.e., a advertising magazine, an advertising booklet, where, in turn, there was advertising from a company that was located in this single-story building. That's why ChatGPT Deep Research was able to find it in the very last corner of the internet, in principle. That's why Deep Research was able to immediately tell me, attention, um, this company was located there around 1970, because that's when the newspaper was from. And that, um, shook me up a bit, because Deep Research searches the entire internet, um, with the speed you know from ChatGPT. And if it then takes 20 minutes, it's not because it's slower, but, um, because it's searching more. And, uh, therefore, this data volume is incredibly large, it can structure and return a lot of data. Everything there is to know about a plot, about its buildability, uh, whether there were ever any political plans, um, and, and, all of this is output. And I'll show you all of this now using, uh, a, I'll call it a global analysis. You see, um, this is quite extensive here, so, uh, what has been output by ChatGPT in this case. Um, how it works, I'll show you how it starts. Let's look at the prompt. How does ChatGPT Deep Research work? You simply start a new chat. I'll set this back to auto again. Enter your prompt. In this case, the prompt is: I want to conduct a comprehensive deep research analysis of a plot. The address is Warschauerstraße 45 in Berlin. Please analyze all available and publicly accessible information about the plot. The goal is to find out whether the plot is suitable for a possible new construction with a residential project. Please conduct the research in a structured manner. Um, this doesn't look as clear as we'd like it to be right now, but in principle, it says: "First, check everything that I need to know in any way." And this prompt also comes from ChatGPT. And now I tell it, now I've made a mistake. Sorry, now I don't tell it to start directly, but I tell it to use Deep Research, and then I click on start. What happens now, um, now it asks me questions again. Sometimes there are four, sometimes there are five, sometimes there's only one. Once I got that it didn't ask me any questions at all. Um, then my prompt must have been very good. Could you please indicate whether you are the owner of the plot or in contact with the owner? Uh, I'm in contact, I'll just lie. M certain information, such as land register extracts, are usually accessible with legitimate interest. Uh, I don't have that. Yes, so it won't find the information anyway. Also, are you planning the supervision yourself? Yes, I'm planning it myself. So, and, um, then you answer these questions once, and then it looks a bit different here than usual. Research is started, and then this deep research begins, which then, and now I'll go back to, uh, what I've prepared here. Um, which then, well, in this case, took 12 minutes, and in 12 minutes, ChatGPT searched 22 sources and performed 86 searches and compiled all the information it found from them. And this comprehensive research on the plot includes plot data and location. Which parcels are they actually? Uh, current use according to the cadastre. Macro location is in which district? Micro location? Um, infrastructure, transport connection is excellent. Um, building law, development plan. Uh, here this development plan applies, it's currently being said. Um, I happen to know that this development plan does not apply, but that it's directly active. That, uh, means with this VO research functionality, um, you have to be a bit careful sometimes, especially with development plans. Uh, that's because these, uh, development plans are on websites that cannot necessarily be searched by a deep research tool. For that, you would use the agent mode of ChatGPT, and I'll get to that in a moment. This information here, um, is otherwise very, very accurate and highly probable, and very, very important. And you can do something with it. Historical and legal information. Yes, what was there before? It was somehow on inner-German border strips. Um, you can, of course, also have the whole thing summarized so that only the questions that are relevant to you are answered, but if you really want a lot of information about a property, about a plot, this deep research, this deep research tool is exactly what you should use. Now let's look a bit at how it works. You can see here, which details it's currently checking. So you can follow it, um, in what it's actually doing now. Yes, I would always recommend, if a plot, a property is relevant to you, to carry out this deep research and to look precisely, um, what I can actually do with it, what, what I can actually build from it, and what the information is that I need, um, for the plot, which I looked at in Düsseldorf, or which I looked at in Düsseldorf, um, that, uh, that, that I, uh, mentioned briefly earlier, um, for example, Deep Research went so far as to say, are there any unexploded bombs from World War II on the plot? No, very likely not, because the subway is right next door. But nevertheless, Deep Research found out that the subway does not run under the plot. Um, so that a multi-story car park would be possible, but it is also very likely that there are no unexploded bombs because the area was searched beforehand. And all these connections, ChatGPT Deep Research puts them together. That means, if you really want to research further, what the real estate agent might not tell you, or, uh, what perhaps no one else has asked about, or what, uh, what no one knows anywhere, then I would always recommend starting with such a deep research, um, and, uh, looking at the whole thing accordingly. Um, let's go into the agent mode, which I've already hinted at. The agent mode is something similar to Deep Research, with the difference that, um, ChatGPT in agent mode can also use websites for you and, uh, as strange as it sounds, can click on things. And this clicking on things, uh, is very, very relevant and very, very crucial, because, um, with it, you can see exactly, uh, what is summarized, perhaps on maps. The example I had just now was development plans. Development plans in Berlin are, uh, on a.
Websites that contain very, very many maps, where you sometimes have to enter addresses to find something, and then someone needs to know how to navigate with maps. And AI, um, can only do this halfway with the agent mode, which constantly looks at the website itself and says: "Okay, I need to click on that now and do this." With the agent mode, however, this is possible. And, uh, for this agent mode, I have brought you two examples. We'll start with the, uh, lower example, because I also want to show you what the whole thing looks like when it works. And the lower example, um, is what we will now enter together here at SHG GPT. Um, it's about a location analysis, and, uh, because it's agent mode, and we'll start the whole thing, but, uh, also briefly touch upon the finished one. There we will see the prompt a bit more clearly, and I will show you what we have actually submitted here. Um, we are now doing a location analysis, a micro-macro situation for residential at this address here. You see that this is no longer the address in Berlin, because I thought that a micro-macro analysis at Warschauerstraße in Berlin is a bit boring. Um, I believe there is, uh, relatively little that, um, that is better than the micro-macro situation, if we, uh, frame and express it generally now. Therefore, I have now simply taken Parkstraße 29 in Osnabrück, um, which is less comparable to Warschauerstraße. And here again, the prompt comes out, or I had the prompt output. Um, validate important facts with at least two sources. Uh, because I always want to know, right, where does it all come from, where does this information come from? Um, and in principle, something like a scorecard is already stored here. So, what are the points that should actually be checked? Learning cards, air quality, um, labor market, yes, uh, urban development, um, image, reputation, uh, these are all things that we, um, that we check here. And then we have this standardized rating system, uh, with an overall judgment for residential from A+ to D. Um, so a scorecard with weighting, right? Micro-location: is everyday accessibility, uh, given, that accounts for 20%, public transport 15%, and so on. Um, I have of course told the AI that created this prompt for me beforehand, please include this scorecard as well, so that I can estimate very, very well myself which areas are particularly good and which are particularly bad. You can of course also build this scorecard yourself, assemble it, and tell me, yes, well, for what I am planning, perhaps nursing homes, nursing homes, the micro-location is the most important and accounts for 90% of everyday accessibility of, uh, uh, which in turn, uh, yes, I don't know, makes up 70% of it, right? Or public transport, then the remaining 30. Um, I can input all of that myself accordingly. And then, and now let's see how the whole thing looks right now, what, uh, yes, what CHBT, uh, Deep Research, the agent mode is doing right now. This agent mode looks a bit different now, and it is now searching live on the websites. Um, you can now see what it is doing, and sometimes it might tell you, beware, you need to intervene briefly here. Um, because if you need to intervene, it usually means that you have to click on things or solve a "I am a human" puzzle, as I would say, so that it can continue working, and it will tell you that, and then it can continue working with it. This also means that this tool is primarily intended to, in the end, uh, yes, take tasks off your hands. What I did recently is I ordered a rental car from Enterprise, and I wanted to have the invoice for it, and I simply, or I just uploaded my booking to Chat GPT with the agent mode and then said, okay, uh, I would now like you to request the invoice online, that's done via a form, and it did that completely automatically. Of course, I had to say who I actually am and what my email address is, but, uh, otherwise it did it completely automatically, and, uh, that's what this agent mode is for, really, to click on certain things again during the search. And what we have now gotten out of this location, uh, analysis. Oh, I opened it, uh, incorrectly, sorry. So, what we have gotten out of the location analysis is, within 5 minutes, a very extensive document, an executive summary. Overall judgment is B+. Overall score, uh, 78 out of 100. Um, because we have full clinic coverage. Uh, away from the main axes, it is generally quiet. We have metro buses, uh, with, uh, neighborhood lines. Everyday life is highly accessible, and, and this entire analysis, how far is the pharmacy at Hoffmeierplatz, 400 to 600 meters. Um, this entire analysis was completely carried out by Church BT alone. The first one or two times you use it, I would recommend you to check again, uh, what comes out of it. Um, I have checked it multiple times, uh, with what I am, uh, what I am doing here. Yes, we don't need this, but I have checked it multiple times, and the information that came out was very, very good for me. Uh, therefore, test it yourself so that you can trust the whole thing for yourself, and then, um, start. Yes, um, the same applies to purchase prices. Uh, there are sometimes websites where you also have maps, where you have to click on certain things, uh, to get these purchase prices. But, um, here too, I simply used a prompt that told me, for Warschauerstraße, I would like the square meter prices at this location from ten different sources. And the square meter prices were then output to me here, each with the statement, yes, we know where this comes from, uh, project page from Imoscout, and so on, and, and, and with that, you can then also continue working. So, that too is, first of all, uh, automated for you. What is important is that this first prompt is always very good. Yes, and that's it for the area of research or the research of, uh, of further, um, of further data. Yes. Um, what we had was extracting data and partly already combining it with a purchase profile, and researching data, and that sets, or, uh, of further information about the, uh, object, the property, what you are currently checking. What we have now as a third point, um, yes, I have already partially covered it in the first point, is, um, matching the exposé with your own purchase profile, but not just using the exposé for that, but also, uh, all further information that you have in certain documents, that you may have also found via Church BT, uh, the reservation mode, or that you are including from any other sources. And what I would absolutely recommend is to use the tool Notebook LM for this. Um, and I'll show you that once. Notebook LM comes from Google, so it works on this Google Gemini. That's the, uh, the language model behind it. And, um, Notebook LM is currently still free to use, you can create a new notebook here, and then you have to upload sources here. What I have prepared is, uh, a purchase profile, um, a, uh, for multi-family houses in Southern Germany, three exposés for multi-family houses in Southern Germany, and, uh, in each case, further documents for these, uh, properties in Southern Germany. This is all fictional for now. Yes, and I'll just upload them here. And you see, right, they are now all here as PDFs. So, somehow here, building encumbrance information in Forze, uh, energy certificate in Stuttgart, here the exposés, land register extract, list of tenants, all fictional, fictional data. What you have here is an AI-based data room. You can upload up to 300 of these documents, and then you enter text here, um, and ask a question or tell the AI to work with it. And the AI will output exactly what you, uh, what you want. That is, if you were to say, compare the three objects with each other, or which of the three objects has less than 500 square meters of living space, um, or, uh, for which object it is clearly that the roof needs to be redone or something. Then the Notebook LM tool will output that immediately. I would strongly recommend you to create a Notebook LM, uh, notebook for every project you are currently checking, for every object, upload everything you have, including perhaps emails, uh, or the content of emails, at least, uh, with the, uh, with the real estate agent, so that you have all the information, and then you can always come back to it, ask a question, and accordingly, uh, extract information from there. And that is what I would absolutely recommend to you. There is also a tool from JGBT for this. That is, uh, um, JGBT Projects, that's what it's called, you could also use that. This one is currently free, and honestly, um, Google is very, very good at this. I would advise you to use this one. There's much more to come. You could theoretically have a podcast created for all the objects now, a bit of playing around, or a video created, or other reports like a learning plan, uh, output. I recommend this for projects, internal and external. Everything where you have multiple documents and data that are constantly being worked on, um, upload them here and ask your questions. In this case, we ask the question: you find a purchase profile in the files, uh, as well as three exposés, um, and in each case, further information about the objects. Uh, compare all objects with the purchase profile and tell me, um, which one best fits my purchase profile, where the risks might lie in the purchase of the object, and what I need to consider. This is, for once, a prompt, uh, that I have, uh, come up with myself. It will take a bit longer now. This is also, of course, already prepared, um, with exactly this prompt and exactly these, uh, exactly these files. Although it could be, right? Yes, they are exactly the files. And, um, well, we still have to wait for it to finish here. But good thing seems to have just been fetched. Um, the old chat here was just, uh, no longer displayed. To compare the objects with your purchase profile and determine the best-fitting object, I will analyze each object and the associated information in detail and investigate the risks and necessary considerations. Um, here all information from each object is extracted, and here the comparison with the purchase profile takes place. Fits optimally. Fits. Fits. Fits. Fits optimally. Financial criteria fits. Does not fit. Does not fit, does not fit, and so on. And you can see here, the answer is also relatively large. This is, first of all, for each individual object. That means, um, what we did in several steps earlier, at one point or another, I would say that it was several steps. This tool can now do this by simply loading something in and entering a prompt, and it can be two sentences long. Uh, and you will find out exactly what you want. Of course, it is still a bit unclear. It would be nice if one could get it automatically in a table. We will look at that shortly, uh, how to achieve that, or at least theoretically, we will look at that shortly, how to achieve that. Um, but extracting this information, that is, first of all, something that is significantly, significantly faster than, uh, than it has been done so far, namely by reading each individual, uh, document. And it's always a bit of a question of how good AI really is, um, with its data, because, uh, hallucinations or hallucinating by AI is known. Anyone among you who has used CHGBT before will also have gotten nonsense out of it at some point. Um, that's how it is sometimes, unfortunately. But, um, when it comes to extracting information from existing, uh, from existing, um, PDFs and existing documents, then AI is, uh, very, very good and very, very fast. Yes, this is something where I would personally definitely trust. And you can also incorporate, uh, funding legislation or something like that here and then say, now please consider everything, or please consider the, uh, energetically worst object in relation to these possible funding options and tell me what I could get out of that. It all works. So Notebook LM, as I said, I believe it is still free at the moment. We are completely on Google Cloud, therefore, uh, it would be free for me too, if it, uh, perhaps is no longer completely free, but this is something where I definitely say, this belongs in every office, um, to work with it in the end. Yes. Um, exactly. And now we will move on to a so-called Super Agent as the next and last point. M, a Super Agent is, the word always sounds a bit, a bit silly to me, honestly, when you think it's completely exaggerated. A Super Agent like Manus here is, uh, a tool that can do, uh, many things simultaneously. Um, it can create images, create slides, uh, websites, table configurations, visualizations, video, audio, playbooks, uh, video audio, and several other things. Um, simply with the input of a single prompt. Why? Um, it, uh, yes, sends individual agents on a journey, uh, for the respective research or to carry out the respective research. Uh, it sends them on a journey, for example, to create an exposé with an AI-generated image. It first has to generate the image, then it has to have the information from deep research, and then it has to, um, put that information into a PDF. Uh, and then you might also say, I also want the following data, uh, structured in a dashboard on a website. But a Super Agent like Manus does all of that by you simply entering what you actually want. And, um, yes, what doesn't work beforehand, let's try it again. Uh, and that's how I, for example, also generated these, um, these purchase profiles and the exposés that you saw earlier, right? Um, because it was output to me very, very quickly. Uh, what I also generated here is a scorecard. So I told it again, look, now please also look at, um, the text, uh, now please also look at the purchase profile and consider, based on that, what is important to me and how, and build a scorecard with which one can then very, very quickly see, um, does the object fit the geographical region? 150 points are awarded for that. So, it just came up with these, these 1000 points itself. 150 points are awarded, uh, if the geographical location perfectly matches what we actually want. And then, for an AI to really understand very well, uh, open this up here, so that an AI can really understand very well, uh, what, um, how well an object really fits the purchase profile, it is important to give it, uh, more information than just the purchase profile. For example, the statement, um, what are the exact target regions and how many points would you give for being in these regions? How good is that and how many points would you give, uh, if, for example, they are smaller towns in Baden-Württemberg, Bavaria, or Hesse with over 50,000 inhabitants, then it would be, for example, fewer points. That means, with this scorecard, you specify exactly, um, how certain assessments are made. And in principle, this is a transfer of what you have in mind, um, how, how you assess certain things, simply to the AI, so that the AI can then go out and, uh, and with its knowledge, with its, um, with its scorecard in mind, assess how good the, uh, how good the object is in terms of geographical location, how good the micro-location infrastructure is, up to the object criteria, um, what are the building sizes? So scenario A would be 3 to 12 residential units, total living area, um, scenario B acceptable sizes and, and building substance, and, and this scorecard, which I have now built here for my fictional purchase profile, fictional, um, I would say, if you say, I am looking for a very specific type of object, um, then I would recommend you to build this scorecard yourself, because probably no one knows as well as you, what the individual scenarios are. Build this scorecard here, then you can use this scorecard again, to then, uh, or with it and with the corresponding, uh, corresponding properties, then go into Manus again and perform this complete, uh, analysis of an individual object with Manus. What do I mean by that? Um, you enter a prompt again. You are an experienced real estate analyst with deep market understanding. Your task is to analyze the purchase profile uploaded by me, the scorecard we just created, uh, for the match between exposé and property and in the purchase profile, as well as other uploaded object files. I've slightly rearranged the sentence just now. Um, you should therefore go through the scorecard in detail, point by point, for every point where information might be missing, research the additional information yourself, and based on all data, provide a well-founded assessment of whether the property is suitable. For this, there are the following rules here, um, which are also included. And what I have uploaded is, of course, the scorecard that we saw earlier. It is displayed a bit differently here now, but okay. Uh, the purchase profiles. Um, I uploaded them as PDF, then they were extracted, and that's why they look a bit different. Don't worry, the information is correct. And what it has done is here, you can also look at what the, what Manus is doing. Um, it has asked me another question, or there was an error, then I can continue. And, uh, now it has checked this one object and says, the purchase is recommended because it has a total of 113 out of 125 points. We need to look once why, uh, yes, okay, it has, uh, it has, uh, it has pulled down the points from these 1000 that we initially had, but, um, in terms of content, it should still fit. It has first done this scorecard analysis and says, for the geographical location, uh, we get, um, 25 points, because we are in scenario A. Uh, there we have 10 points, fits. For micro-location, we are also in scenario A, and it is justified, fits too. And for infrastructure and transport connections, we are in scenario A. Um, so the best, uh, that we actually had in there, also gives, so in this case, five points, because it is simply, uh, weighted amongst each other. And, uh, you could continue this way through the object criteria. I'll just look once, uh, financial criteria, uh, no, it has also been able to calculate a bit here, it has calculated the purchase price factor. That was not in the exposé, in my opinion. Profitability indicators, uh, it has also calculated. So, if you want such things calculated, uh, then you must always tell the AI exactly beforehand what should be calculated and how it should be calculated. Yes, financing, uh, tenant structure, uh, current rental situation, and so on. So all of this is automatically checked, entered into this scorecard, and then on one hand, you get the scorecard, and on the other hand, an executive summary, and you see, does it make sense or not? In this case, there would be a negative cash flow, uh, calculated. Um, yes, uh, that is, and the weakness is a high purchase price factor, 22 moderate returns. What you can also incorporate here, research, right? What you can also incorporate, uh, are these complete micro-macro analyses that we already did earlier. And then Manus is very, very good at providing you, uh, within, I would say, 20 minutes, a complete research, uh, on this object and outputting everything that, um, you might possibly need to know, including, for example, bus connections. And all of this is, it's just one prompt away, I would say. Yes, it's relatively fast. Um, and oh, if you look at the knowledge suggestions here, a small additional tip, uh, what is output here, then Manus will understand what you want if you only upload the exposés. Uh, and then you don't need to enter the prompt anymore. That is, once done, right? Upload the exposés, and then something will come out here accordingly. What comes out here, you can also have it output as Excel or as anything else. So you have super many possibilities, also as a website, as a landing page. It will be, uh, created for you, uh, at the push of a button. Um, and through that, that was the last point, using Manus for global examination of all relevant criteria. Now I want to show you one more thing at the end. This is one that, uh, this is what I announced earlier, which is slightly theoretical. What I did is, I believe this is now another scorecard, uh, another scorecard output. Um, namely here with the, uh, with the information, if the location is in Berlin, then you get three points. If the location is in the S-Bahn commuter belt, two, and in peripheral areas, one point. And, um, this scorecard you will now find here in the overview with the corresponding relevant points that are relevant to me when checking the, uh, exposé. And, uh, now you already see two, um, example addresses that are included here. Um, they were checked accordingly, yes. So here we only have two or multiple two points, but mostly three points. Um, this exposé here was checked, has a total score of, uh, 57 out of 60 and thus matches 95%, thus excellently matches what we are looking for. And how does this get into this Excel table? I used a tool like, for example, make.com. There is also Zapier, um, or N8N, N8N is what they are called. Um, these tools are for connecting different tools. What have I connected here? First, my Gmail account. Um, and it knows every time, from this email address, I could also build it differently, every time, from any email address, with an attachment, but, uh, if I receive an email from this email address with an, um, with an exposé as an attachment, then this exposé is automatically uploaded to LM as a PDF document, and then compared here with my purchase profile and with how it is entered into this scorecard, and at the end, this LM provides a result, and that is automatically entered into the table. It may sound like some of you have already switched off, if you say, okay, that's a bit too much technology for me. I said at the beginning, I am not a, um, a techie, not a developer, I can still do it, and I can do it because I simply ask the AI how to do it, if I simply talk to CHGPT and say, I want to build such an automation via Makcom or something else. Um, most of what is written here, I have no idea about anyway. Uh, but I simply ask the AI, uh, put in 242 screenshots here, if it doesn't work, and say, what do I need to do now? And then the AI builds it for me, or the AI tells me exactly what to do. And that's how I build such a process. And this process is, of course, relatively simple now. This is now, first of all, an extraction, uh, of the data from the exposé with Cloud, with the other Large Language Model into this table here. Um, it's not yet that it's being researched somewhere or anything else, but you can build all of that in. This whole thing here is so freely, uh, buildable, I would say, that you could now, for example, incorporate Chat GPT here and have Chat GPT output further information, um, yes, uh, which would then also be included in the table. Start. So, it's step three for most of you, I understand, I know. If you want to do it this way, then start by thinking of it very simply. An email comes in with an attachment, it is sorted, and then it is output here in the table accordingly, and then you can gradually add more things, which you are currently still checking manually, one by one. And then at the end, we have such a scorecard for all properties that come in automatically, and you know, um, yes, if this is somehow over 80% hit rate, then I will look at it myself. Otherwise, I will no longer look at it myself manually and oops, uh, myself manually and, uh, and manually, but only if it's above a certain point. And then you could extend this automation so that the email is automatically answered, sorry, it's not for us. Yes, all of that is very, uh, very easy, and it works via so-called interfaces and APIs. That was now the, uh, last part, which really went very deep. Um, and now I will look at the, uh, questions again and will try to, uh, pick out the, uh, most important questions. What do you mean by most important questions? I wouldn't evaluate it like that at all, but the, uh, the ones that have perhaps been asked most frequently. Um, you will receive the recording, as in the past, for these, uh, webinars, it will be sent to you as a link. All our webinars can be found on YouTube, free of charge, which you can watch again and again. So also the last webinar that I held, for example, on the topic of, uh, yes, one hour of Chat to BT for the real estate world. Uh, and before that, we looked at 7, 8, I don't know anymore, uh, AI tools and what, what basis they have in the real estate world. Um, there was also a question about data protection, technical and legal data protection aspects of these Custom GPTs, whether that is unproblematic. Uh, it always depends on which, uh, which form and which type you use of Chet GPT, and of the Custom GPTs. Uh, so one says, it is always, it is always a very difficult topic with data protection. Uh, it is very often said, yes, we can't do that at all because it's not possible for us from a data protection perspective. Um, first of all, the general chat usage anyway, and that is honestly wrong. Um, from a GDPR technical perspective, it is possible to use Chat GPT. There are also tools, these problems that you have, they will, um, they will be faced by many others as well. Therefore, tools are also being built that take away exactly this problem for you. So to say, I won't even start with it because there could be, uh, data protection problems, uh, that is the worst, because your competitors will start using it, and then you will have a problem at some point. Um, I would advise you to solve these problems, uh, that you, that you see there, um, and then start directly. And the problems are not as big as you think. Um, I would claim at this point. Otherwise, regarding the specific question, I don't want to go into detail now because, uh, that would exceed the scope a bit right now. We still have the question whether we can upload a 360° tour with Gemini. To my knowledge, that is not possible. Um, then there is also the question whether I can provide the prompts used. Um, I will see, uh, that I, uh, perhaps make it available somewhere, I didn't do that last time. Um, Deep Research does not seem to be available to everyone. Uh, Deep Research is actually available to everyone. Um, you have to buy the 23 Euro per month version. Uh, but it is definitely worth it. So, I think you will find that out very quickly, and then click on Deep Research, and then it will work. Yes, um, that's possible. Um, can the GPT search function also be run via Side? Very good, uh, very good question. Um, it's not possible via Chat GBT right now. Via, um, via Plaud it would work via another tool, via a so-called NCP. We can talk about that sometime. We could, uh, gladly show you how it works sometime. Um, we have already built it, man. Whether chat GBT can use the agent in agent mode, I cannot say 100% right now, I must admit, because, uh, I haven't tried that yet. Yes. Um, the further questions, I still have them. I will answer them by email because I have just seen that I am actually looking at most of the questions here. Um, and would at this point, with a slight overrun, uh, let you go into your lunch break. Um, if you want my contact details, because you have questions, because you have ideas, then please scan this QR code. Uh, if you enter your email address there, you will receive my contact details. That is the only functionality, uh, that this, uh, that this has. Um, you are very welcome to do that. And, um, then on Friday, we have a second webinar. Uh, it will be about how you can, with the help of Jet GPT and Side, uh, yes, create project developments within 30 minutes, uh, so that you have a very good basis for decision-making. Uh, or honestly, it goes beyond a basis for decision-making. Uh, what you can create with Chat GPT and Side within 30 minutes is, um, immense. And therefore, I would recommend you to also watch the seminar webinar on Friday with my colleague Annika Meisters and, uh, our CEO Matthias ZKE. Um, I think if you are interested in this, you are also interested in the webinar on Friday. Um, yes, registration for that works via this QR code. You are also welcome to do that. Yeah.