Myth Busting The Impact of AI on Real Estate Valuations

Welcome to the latest edition of the AllSop Prop Chat. My name is George Walker and today I've got the great pleasure of mediating a podcast where we're going to discuss AI, myth-busting the impact of AI in real estate valuations. Today I've got three knowledgeable people who are joining me. We have said Verity, head of research and data at AllSops, Freya Murrah, associate in the commercial valuations team and a colleague, Evie Robinson, who's desperate to get a win-win, but she's squeezed in the softening chat for us. So thank you very much and welcome everybody. Thank you George. Thank you. Okay, so to start the podcast proper, let's just introduce folks around the table. So Freya, give us the quick rundown on Freya Murrah. So I work in the commercial valuation team at AllSop, as you mentioned. I've been in industry sort of just over 12 years now, a qualified valuer for seven and I value a range of commercial assets, primarily for bank secured lending, but could vary from there, from accounts, internal reporting, that sort of thing. So that's pretty much my day today. Brilliant, part of quite a big team now, aren't you? We are indeed, yeah. And we're national as well, so we've got coverage all over the UK. Brilliant. Okay, so Seb. Hello. Not a valuer. No. A data man or a search man. Yeah. It was the update. What is the update? Yeah, I've got about three hats still. I've did presentations more into a bank. That's one of the hats. We've got a client, a client data reports job on the moment as well, consultancy reports, that's what hat to write all the market conditions and try and keep a handle on everything that's going on until everyone what the market's doing. So yeah, that's that number three research and data and a very broad road to practice like AllSops has got there has got many hats. And so is there a resie bias to what you do now? Or is it you taking in bleeding into commercial attack? Yeah, I would say yeah, there still is there's a lot to do in Resie and a lot going on. Yeah, a lot of policy change in Resie, which has given me a business. Go good. So, Evie, the quick buyer. I've been in the valuation team since 2020, the commercial team. Been at AllSop since 2018, doing my grad scheme, working alongside Freya and our wider team and focus yet on anything commercial and national. Yeah, that's classical AllSop. You guys aren't really sexist specifically, are you? No, you are agnostic in that regard. You are agnostic, which is a problem. Okay, so today, I mean, what we're trying to do is, where is AI with commercial valuations? Has it impacted the world? Myth busting, has it had a major impact? So, I think let's start if we can get said, we've all tried to define what AI is and you are the shortest and most concise. So, give us what your interpretation of AI, the definition for AI. I'll do my best. It's essentially a mechanism through which human thought processes are mimicked. And at the moment, that means fairly linear processes can be automated to a degree, rather than any creative or, yeah, innovative creation. It's automation of a linearish thought process. Processes. Okay, so it's not yet predicting in this context. Not in a very creative way. It's still quite linear in it, what can be done, but that might change in the future. Okay. And it's interesting, we'll go on to the regulation later, but the RSS have had a crack at very recently, defining how it should be used in the profession, which is, of course, essential, but let's come on to it later. So, where are we, what systems for the moments are in regular use, which have a bit of AI involved in the valuation process? To think about valuation explicitly in terms of literally coming up with an opinion of value. There's actually not that much that we regularly use. I think historically, people will know about AVM, so automated valuation models. So, this is where data is pulled and said, correct me if I'm wrong here, but land registry data in a residential context and is pulled, and then it'll spit out a price most likely for residential assets rather than commercial. And that is adopted in the market. I don't think necessarily by professionals, but certainly by investors or property owners, residential owners, that sort of thing. But other than that, in terms of actually a valuation model that we would use as professionals that explicitly uses AI, it's not widely adopted. Okay, but you've got lots of tools, I mean, you mentioned Valos and things. We've got several tools, things like Kel and Valos and Argus, to name a few of them, but they are assisting us rather than actually doing valuation. And I think the key distinction there is obviously, yes, obviously, it's about the inputs that we are putting into that data platform and pulling it out of it. I think the difference between sort of AI in terms of helping us do our jobs or automating what we do, it's being able to pull that out and then use that. Whereas at the moment, it is still very labor-intensive on a valuer to make sure that that information is correct. Whereas the direction of travel where we hope to be, whereas we become more reliant on those platforms that we do use. So Evie just, well, and George, you mentioned there, and it's not necessarily an AI platform, but it's a form of AI that we've looked at doing, which is Valos, which is to do with the way in which that we pull data for our reporting. So there's a lot that goes into valuation reports. So obviously, there's the figure that comes out at the end, but in terms of RICS compliant reports, there's a lot of data that is factual. Which make up 50 pages. Exactly. So that is the probably the most fundamental and important thing that we as valuers are focusing on at the moment is automating that factual data, obviously still while checking it and verifying it, but being able to pull it efficiently and quickly. That data includes things like planning rates, any government websites that we might use, council tax, things like that. So it's the things that the valuer will spend quite a lot of time, maybe hours, looking through each document, each planning document, each council tax to see if we can locate what council tax or what rateable value something is, if it's not easily sort of findable. And that's what's helping us save time and become more efficient using Valos, which has been really helpful. So you're going to come in. Can I be a stickler? Go on. So there's two things going on here. There is AI in our industry and in your sectors, but there's a very gray area between what is actual AI and what is process automation. So a lot of that sort of fetching of information is process automation in a significant degree, but then the NAI's job, if it's used, is to make a judgment call, that sort of human reasoning bit, because they hang on and it is this bit of information that's being extracted, relevant, useful, appropriate, and so on. And then pattern, if it's generative, package that into a block of text or something to then render it to screen. Give us an example of that, so to make it clearer. Take the last example. Okay. So yes, you're right. It pulls planning information records. It creates and technically speaking, it searches and renders those planning records to screen. There may well be an element of judgment and reasoning attached to that process, but the simple sort of fetch based upon a geographical or an address sort of match and so on. There's no AI in that. And a lot of the AVMs, historically anyway, aren't, they're not AI based and I don't even have a home tracker using AI yet, who are the main kind of lenders, Resiavia. It is a fairly complicated traditional data science and modeling where AI is increasingly in uses. It's in that adding reasoning to the results of those fetches. That makes sense. Classy example, I know people using it, in the valos and the location context, is things like what the location description, which is an element of making a judgment about where this thing is, how close is it to X, Y and Z, and creating a paragraph and generic vanilla paragraph about it. But it's pretty simple stuff, a little bit copy and pasty. But it's a brain dead task for some poor surveyor or grad who has to sit there and type out that this property is located in the vest of Jumsford and his ex, their users for whatever. It's what I remember for doing my old, I'm not a valio, but I did my grad stint as a grad surveyor and that's always the thing. Did you last? I did about a year or so, yeah, and then I got shunted into research. I got too good at hunting for comps, it turned out. But yeah, it's one of those things that you have to write. It's almost your write-or-path, write-and-wise location descriptions. So on the subject of the low level work, I mean KPMG are famously 10% less graduates this year. I mean, are we seeing less people of vulvas industry? When I go past your team, which I do on a regular basis, I just see a lot of huddling going on. And you are constantly chatting to each other about what is actually, as you know, in 90 seconds, I managed to value Acura's house on their chat GPT today. And when you think of evaluation reports, 50 pages, this was about 17 lines and it took me 90 seconds. But I see a lot of collaboration in your team. Are you seeing it replacing bodies on the ground or how is that impacting numbers? And the man time needed for evaluation, a proper valuation report. So for us also, no, I think is the sort of the short answer, but the slightly longer answer, I would say is because it's still very much that human touch and that interpretation of that. I think it's been fascinating for me in terms of where I sit in my career. So this wasn't necessarily something that was available when I started, but it's now something that's being regularly used. But obviously the next generation down, for them, it's something that they feel quite comfortable accessing or using and obviously we'll go on to talk about the sort of the risks associated with that. But they are much more agile at understanding this and they can become much more efficient and work much more quickly, and therefore ultimately become much more valuable, much faster. Is that fair to say? And I think the use of certain AI platforms, I'm sure everyone's aware of, they are pulling information that they might have had to ask three different colleagues for or they are doing sort of finding equations that they might not have quite yet been taught at university, that they are then able to implement. And we had an instance the other day where one of our apprentices was compounding a rent and she put it all into the database and it popped out with a number on an AI platform. And I said, that's great, but can we just check it against the formula? And we both checked it and we got another colleague to check it as well. And the actual figure that popped out was incorrect, but the equation was right. So it was a very helpful tool to use, but equally, we had to double check it. In my world of binary data, that should be possible. So surely you can. Because all data is a one or a zero, how do you get the wrong number? Yes, but it depends what depends what the LLM or whatever the model that whatever the AI model is that's sitting behind that screen or that agent or whatever is doing with it. And that's part of the problem around transparency as to unlike a human colleague, you can't sit over its shoulder and watch it type formulae into Excel and come up with that number because it just doesn't it's just fired up today to send her back you come. Okay, what have you done with it on that journey? And do I know you've gone to the right place to put something up? Do I know you have, in fact, put your brackets in the right place in your Excel to not to not have to hallucinate the wrong number or in this case, hallucinate an incorrect result. And we've tested this ourselves it also with some data extraction AI tool, genuine AI, scanning through lots of documents, trying to pass them and make sure that they're looking for certain key terms in, let's say a lease, for example, and structure that into a pro forma. Now, at the moment, none of the LLMs or other models that we've had that try and do that come back with sufficiently good enough results to make you not want to check all of its results. And that's kind of the threshold, say, well, fine, got the commencement date, let's say you've got the term and you've got parties, fine. But hang on a minute, you've missed clause A, B and C that we were looking for always, or you've kind of not quite got that right, in which case we have to check everything. And is that because you're not asking the right questions? So my next question is, is the skill of the value changed in actually asking the right questions of the technology in front of you? I've been saying to people who've come to me and said, can we get an AI tool? Should we do this? That is well, it comes kind of down to prompting. And I think you still have to treat it a bit like a grad, have to point them towards the right information, tell them what they're looking for, check the output where they've done it, and then maybe slightly redo it if needs be. I don't think any of that changes with AI as it would, let's say a new grad on the team, you're trying to walk them through something for the first time. And then there's also the value of an experienced valuer is a suck it and see, does that seem right? So one component was clearly obviously wrong to you. Then you get the finished product, is that number correct? Again, that's another skill we all have. I also think with leases particularly, often they can be very lengthy documents, and that AI tool is helpful to say, this is on this page, or search for it here, it's in this clause, and then draw your attention to that page. I think that is actually incredibly efficient and a really good way of saving time, because often leases are written in different ways. But you wouldn't want to ground to necessarily read a lease and come up with a reason set of advice for a client straight out fresh out of the gate, the same way you might want them though, to just can you get the highlighter and show me where I'm not the lawyers. Do we know what the lawyers are doing? Have they managed to correct this? Investigate leases? There are now legally targeting LLMs and so on. Yeah, the AI tools that are effectively I'll try that, but I think a lot of it is that slightly reasoned process automation. Like you guys, we say Valos, you are back in the day, what 15, 20 years ago, we were talking about mail mergers, like you'd have your word doc and you'd have an Excel sheet and you'd be trying to join the two up. It's not too dissimilar now, it's attempting to just shave off those document production bits that don't make any money or are very boring for you. I think the lawyers are doing the same sort of thing. For them, they've got that great sort of lookup problem, getting to legislation.gov or case law and trying to find, so kind of like taking a legal Google to the next level it would see what they would do be doing. But again, I think they would still be struggling with the same things we would struggle with professionally. And what's the next stage? The valuer is now someone who is doing the same thing, which is taking a lot of information, does that seem sense, taking where ever it's come from? What's the next stage of AI helping that? I think a lot of people think that valuer will be out of the job after, I don't know, it progresses even more. But I think at the end of the day, property is a human business with people's business. You still need the people to look, make the judgment, review different comparables, review the actual asset itself. No asset, particularly in commercial, is identical to the other one. And that is the crucial thing for our side of the business. And like you've said, you can value our new building on AI. And that comes up, pops out with an answer. And I did it with my flat the other day, and I thought, well, it's just pulled lamerage details, but it hasn't said that we've painted it or we've refurbed it or whatever. It's no judgment. And therefore, yes, it's going to help us a lot. It's going to become a lot more efficient, but it's not going to put us out of a job because we still need a human to look at it and say, right, market conditions suggest this. Occupiers have changed their way of occupying a building. So we need to reflect that this building is no longer fit for purpose, or x, y, and z. So I think that's probably where AI is going to, is assumed that it's going to, going to take over in our world. I don't know if you agree, Fred, but. No, absolutely. I mean, it's funny, we were talking in the office the other day about analyzing comparables, because ultimately, I suppose you'd argue, well, it's factual data, it's a price, it's a yield. But the way in which you can even interpret a comparable could be different compared to value, to value, is it topped up if you have made a different assumption about the break laws and things like that. So that, that human interaction will, in my opinion, never change. And obviously, George, you know, your background in terms of buying and selling, you're not going to buy and sell from a computer, you know, you need to have that, does it? Well, Seb, you're making a face. What I mean is, I think you need to have that, but does it sound right to you that you need that trust? You need that information of speaking to someone that you trust? And the head of the credit committee, any of your clients is going to want to be able to ring one of you up and say, hang on, you've done five evaluations, this seems a bit odd for what reason. I mean, will, will our clients running basic programs across what you produce for them? Have we got to that stage yet where you're. Yes, I had a conversation with a client of ours last week, and he said that he puts our reports through their specialist AI system that keeps everything confidential, but they put, they run it through and they get the summary of what we have said, positive negatives, value, why it's this, why it's that. And he gets a one page out of our 50 page document and that, that is, that is how he does it. And then he can report back to credit on it. So yes, they are using them. That's a great use of it. That is a very good use of it. Great day information for me being fresh. I always say that we're in the business of reassurance, no matter what we do, but everything we do, every sort of function we provide involves reassurance and research termed it's providing reassurance around data and you have the fact, correct amount of facts to go and do the thing you're doing. You guys, it is, you are reassuring the client or the bank or whatever, that that is worth that. It's a party on that and the sales perspective, people who are spending money on a serious asset like to hold someone's hand as they walk in the door and make an offer, it's reassuring to have someone backing you. Yeah, until we, unless we get to a point where people, as you say, are reassured by a machine to that's the same extent the demands on, on us using it and relying on I think won't be there. But it'd be interesting for you, your point about the younger crew, I think there's someone younger news is wonderful, but it's their norm and so they, my lot, my next year, English walkers rely on AI and then they build, they go back there, that's the solution and they prove it. It is another way of looking at it, isn't it? No, of course, I'm bringing my own personal view and my own bias, I suppose, because that's another thing as well. Obviously, we have unconscious biases as humans, but you tell me, yeah, in terms of AI platform, surely they've got integral bias as well. Based upon the training data that they've been given. Yeah. You know, you'll get a different set of outputs. You go to GROC to chat GPT, OpenAI or DeepSeq, they will get the same prompt, you'll get different answers, because you say, yeah, they've got different trainings, different biases, different algorithms behind them. And again, comes back to transparency, can't see the formula that's ended up at the value of, or inaccurate value of, accurate house, necessarily. That was quite funny really. Yeah. So in practice, and we've, so we've seen, we're sort of slight nibbling away, but using it, you guys are using it to gather information, you can do what you filter. But any actual practical examples where this model of using AI a bit too heavily is sort of come back to haunt people? I think there has definitely been some of that. I know that, like you've said earlier, a lot of grads are being, intakes are being reduced and things like that. I think overseas sort of help is, is probably being pulled in and I think, I mean, not, not London centric, not London centric, and I think for valuation, that is, you have to have an area where you specialize in and whether that's the UK, Europe, wherever. But I think we have to know the market specifically when we're valuing and you have to have an opinion on the market. And if you're taking that overseas to somewhere where they might not even be seeing the building, they might not see the local data, overseas banks valuing UK property, you're talking about outsourced teams in some far off land. Well, the AI itself is overseas in many cases, it could be. So, ultimately, going back to the risk guidance and so on, ultimately, where is this thing that is giving an opinion? Well, it's not the strongest concern, it's nowhere. Okay, so that have teams been built away from the geography of the property and they're trying to get it right with AI and perhaps you think that's... Yeah, and they're writing locations and descriptions of buildings and that's fine. Location wise, you, again, need to have inspected the building to see what the retail pitch is or whether the shopping center is completely empty or whether it's busy. And I think that is pretty crucial in our valuation reports. And I think that being outsourced is probably going to have an impact on the quality of work and how a report is finished and the level of detail as well. Maybe cheaper, but actually... Definitely cheaper, but I'm not sure... It's going to do the right thing. ...quality will be picked up. Yeah, in the auction world, is it trading and who's trading there? It's a sort of pretty race principle. Yeah, well, I think the question of cost is actually really important in terms of where things maybe have been a challenge. So, say, for example, the process that we have been through to use Valos as an example, that's been an onboarding process of at least six to 12 months because we've been building what we now use with them collaboratively. It's not an off-the-shelf product necessarily. It's very much an all-stop product that we have used their platform. So, I've obviously heard stories of other firms who have potentially gone on their own and built their own platforms or using other AI platforms that maybe haven't worked. So, if you think about the cost implication, not only financially, but time-wise, obviously... So, we're sat around here doing a podcast, but this is now pulling us away from our sort of the earning capacity. I mean, it's important, and we enjoy doing those sorts of things, but making sure that you invest in the right product and in the right data and in the right area, it's a real business risk. And I think that's the next phase of making sure that we understand what's important and what we really want to get out of any platform that we invest in. I think particularly the all-stop report is apparently the most difficult one they've been dealing with, because it's got so much detail and so much information and so many different ways to analyse things. So, I think that's been pretty... Well, I feel very proud of that. Yeah. The detail you go to. Is that right? And very much, testament to Freya and another one of our colleagues who have been running it, well, several of our colleagues have been running it. But, yeah, it's a pretty detailed report, and they have been pretty shocked by how much detail we go into, in comparison to other bits involved. It's a good shock. Well, I think that you should be proud. We see it as this is the start of something that we are going to use going forward for years. So, the hard work is now to then reap the benefits going forward. So, how long has the process taken? Well, as I say, it's been about sort of six to twelve months, and it's ongoing. Just from our perspective, Freya and several others have a weekly call for at least half an hour to an hour every single week and have done for probably six months. Yeah. Yeah. Yeah, I didn't know that. Well, okay. That's what's required to get you things right. Absolutely. So, we're actually saying that that's not pure AI, but these systems are great, but it takes a lot of time and effort and expertise to get them performing to how you'd like to perform. It's quite interesting, isn't it? Yeah. Brilliant. Gosh, that's probably... So, we mentioned in the RSESA their consultation finished in April, didn't it? And they will, I'm sure, come back in due course with a guidance document, which is final. That's right. I think we're anticipating around sort of the autumn time that that will be finalised and sort of circulated, because I think the RSES were probably getting a lot of pressure from the industry. They have a duty of care to advise their members on the appropriate way to use AI, because we need to use it, but it's how we use it. Well, the draft guidance says you must use it, which I was quite delighted about when I read it. There have been many RSES things, but I thought, well, that's actually quite innovative. Yeah. We have a real responsibility to make sure that we do use it responsibly, and I think that will be the real tone and the cadence of that guidance when it does come out, having sort of had a look at the draft guidance already. It's things like transparency. So, making sure our clients are aware that we're using these platforms and these products. Another is the reliability of AI. We've already talked about that. So, can we confirm where the data is coming from? Are there any inherent biases? Things like that, but also privacy and confidentiality, and if I was sitting here as one of AllSOP's clients, that's probably going to be up by my risk curve in terms of, well, if they're harnessing this data, where is it going? Is it held within AllSOP? Is it more public? Obviously, the RSES need to give guidance on that, and it will be following all the sort of normal rules of conduct that we will follow day to day anyway, such as transparency, honesty, integrity. Which is why we don't blast these things on chat GPT or any other open system. Exactly. Fundamentally, we're keeping data and our clients safe, and that's what we are going to do. And it's generative AI. So, it's not predictive. So, it's not going into our data, playing around with it, moving it somewhere else, doing anything predictive. It's a language model. It's waiting things up. So, conclusions to that useful chat. I mean, it seems that we're on a journey already. We're using, as you'd expect, the tools that are available. Is there going to be a massive shift in the next two years, Seben? Your view of AI itself is an exponential curve of innovation. Full-fold increase in revenue, apparently, in the next two years, but that's a lot. Yeah. And capabilities are linked to chipsets and data-centered capacities and so on. So, it will get better and different in the future. The question, I guess, is, to what extent do those kind of human thought processes that it's mimicking shift from being linear, being more erratic, more creative and unpredictable, and slightly more spontaneous or more easily passing the Turing test of this was not written by a computer, it was definitely human. Oh, hang on. It was AI. It's how much further will that envelope be pushed that actually it can genuinely start doing human replacement jobs where there is an element of experience, judgment, and assessment required that are then into proper human place. It was somewhere off that. I think we probably are. In a search, do Google the Turing test because it's quite fun. It's great fun. Yeah. It's quite quite stimulating. Yeah. I mean, it's not going to replace me yet, I hope. So, it's reversible. But data science, for example, there's a degree of replacement going on there. What would have been traditional data modeling, say, 10 years ago? It's significantly faster now in AI and it's much easier for somebody who is a non-data scientist to prompt that type of task. I guess prompt jockeys. Well, this is it. Prompt engineering becomes now a job and is a significant role. But then you're into the questions around liability. Is the prompt engineer liable for a poor prompt and a poor result? Is it accurate? So, as a valuer or the valuer is using an AI system, is he liable for the poor prompt setup daily to check and then publishing presumes? Well, you could ask about the when you're calling the comps. Did that include a rent free in a basic level? Before we leave it, have you got more exposure to residential than the rest of us have? Is the impact different in residential? Probably a lot more transactions. Yeah. Scale is different. Scale is much larger. Data quality is poorer and there's more opacity. So, I mean, you talk about land registry, price paid data, fine. That gives you the price and the date. There's no information anywhere really that's robust around physical characteristics of building. It's an address. Yeah, it's an address and it's a price. You don't know what it is. How big it is, honey bedrooms condition, none of this. Even planning histories at night may not get hold of. So, any model that's trained to be like a manual model or an automated model is struggling with poor quality of data in residential. In which case you put poor data in and you get poor data out. In the most commercial worlds, you've got a VOA statement of area whether you believe it or not. Ask Mr. Boucher. He'll tell you not to. It's more of us than ABC. It's there as well than ABC. So, Resi has a lot more way to go for AI to be robust enough in an evaluation sense or even just in a generative sort of copy sense. Give me some content around values in their SIDCUP over the last three years. Well, again, there's not really sources for it to go out and get there that are particularly robust. So, I'm fairly confident we aren't going to under threat for that in an immediate short term. Same belief, Evie? Looking for the next two or three years? Yes, I don't think we're under threat. I think if anything, it'll just help us be more efficient and make the most of our time and get better quality work for our clients, be out seeing our clients more and actually just have a better relationship and quality of work. That would be my view on that. I think it's only going to help us as valuers. Would you be as bold to say there'd be better, better, more accurate valuation as a result? Yes, because I think you have more time to actually think about the numbers rather than pulling the location and description. So, yes. This is the bonus. It's freeing up time from unvaluable tasks, so you can vote for unvaluable tasks. But for me, AI does help scanning through, let's say, a Bank of England monetary policy report, which can be 100 pages. And I can think about it, so I can get a summary really quickly at 100 pages or generating a summary thereof and then publishing it around you guys. And you can let all the information sort of settle in the mind and think, well, when you come get your things signed off with us, well, for these reasons, I think it's X. Absolutely. I don't think it's a long-trained journey. Yes, so Fred, your final thoughts. So, one thing we haven't mentioned at all today is ESG, which is something that I'm particularly passionate about. And I think echoing what Evie's just said about how it's helping us, something in the commercial world that we are going to be particularly keen to focus on is energy efficiency of buildings and how we harness that data in order to provide more accurate valuations and advice to our clients. And there's a lot of data. Absolutely. So, my hope going forward the next couple of years is how advancements in the harnessing of not only EPC, but energy efficiency of buildings and how that will in reality affect the value of assets, so in a real world scenario. So, cost of occupation, as well as rent, as well as rates, yeah. And EPCs are only going to get you more complicated, more detailed. Do you have a bit of an EPC day? Have you got an EPC day? Don't set Seb off all EPCs, we'll be here for 45 minutes. Well, thanks everybody. That's very insightful. It's an interesting world. It's not the final solution, quite clearly. I hope we've busted a few myths. Thank you very much for listening. And we look forward to hearing you, talking to you again on the old SOC project. That's goodbye from all of us. Thank you. - Thank you. - Thank you.