Ep 270 – Can ChatGPT Do Your Property Development Feasibility? (We Tested It Live!)

Introduction

Hilary: So Aaron, hand on heart, did the robot get it right?

Aaron: It got it very confidently. It was very impressive, but no, it did not get it right.

Hilary: Hello and welcome to the Property Mastermind Podcast, episode 270. Today, AI versus the developer. We covered AI a few weeks back and more than 2,000 people watched that episode, so we thought it was time we did another. You are going to learn a ton, and you are going to know that you have to jump on the bandwagon, but you can’t always trust it.

Hilary: Today I’m here with Aaron Ralph, business partner in Ultra Urban, our development company, part of the Property Mastermind team and an absolute AI guru. We have been having a bit of a giggle about AI and feasibility, so we thought we would unpack it. After the last episode, where so many people watched, we thought it was time we showed people how good it is and how dangerous it can be.

How AI has changed Aaron’s working day

Hilary: For anyone who missed the last episode, the takeaway was really that the barrier is knowing what to ask. How much of your day is used by AI?

Aaron: It’s funny, I was chatting to a mate of mine today who is a lawyer, and he asked how my workflow has changed. Within the last six months it is unrecognisable. My workday is completely unrecognisable, and my working team is 24 hours a day, seven days a week now. My role has gone from completely hands-on to basically orchestrating. The models are more than capable of doing that now.

Hilary: When you say that, it doesn’t mean you’re working seven days. Do you find that having AI around you gives you flexibility?

Aaron: Correct. When I say that, I mean the system itself. Like I mentioned on the previous podcast, this is truly the way business is heading. If you’re not building something that sits on a foundation where it can run 24/7, you’re going to get steamrolled by companies that do.

Hilary: I look at my day and I am so much happier. It has made my life so much easier. I’ve got flexibility, I’ve got an AI team doing things for me, and processes set up. My agents know: when that happens, go do that. You’re no longer having to do it yourself, or getting somebody else to do it who could make a mistake, or who is sick, or can’t make it. It doesn’t fall back on you anymore.

Aaron: That’s right. It’s the concept of cognitive load. Pre-AI, a lot of our attention got used up because we get tired over the working day. They always say, if you go to prison, hopefully you don’t see the judge at the end of the day, because statistically people end up going to prison more because of decision fatigue. Where AI can really stand in is picking up the low-level stuff that isn’t what you want to be focusing on, particularly when you’re doing development or building a business. Those tedious tasks are the very first things you notice you can get rid of, which frees you up to focus on the things you really want to be doing.

Hilary: Like golf. Or running up mountains. We always say we need more AI because Aaron wants to run up more mountains, I want to play more golf, and Bob wants to do more fishing.

Today’s topic: feasibilities, the numbers, and when you can trust AI

Hilary: You and I both love AI. It has changed our lives. My AI talks to Aaron’s AI. I can send mine a message to tell Aaron something, and his sends me things. But today we’re going to talk about feasibilities, the numbers, and when you can and can’t trust AI. We’re going to screen share, so if you are listening, you may want to head over to YouTube and watch it, because Aaron is going to show you the prompting he does and what comes up. We will talk through it as much as possible, so if you’re driving you will still understand what happens, but this one is definitely worth watching. So let’s start, Aaron. What are we going to do?

Aaron: I thought we’d take a look at feasibilities, because what jumps out at me, and I get asked this a lot in the mentoring program, is how helpful can AI be with a feasibility? The answer is it can be incredibly helpful, but you still need to know what you’re doing. You still need to know what you’re looking at.

Aaron: Development is something where you really need that 20,000-foot view. It is one of the rare careers that rewards you for being a bit of a jack of all trades. Throughout the industrial age people became specialists and got rewarded for it, but in real estate development we really are jacks of all trades, so we need to be fairly confident in many areas. We have consultants, of course, and this is where AI can help fill in some of the gaps. But the most important thing is the numbers. At the end of the day, as a developer, we must take responsibility for the numbers.

Aaron: That’s why this is such an important topic, and why I thought we’d do a demonstration to show how impressive AI is, and yet how off the mark it can still be while appearing very confident. Because it has been fed a whole bunch of information, if you’re not hyper specific and tailoring it to your situation, you’ll find it pushes towards the median. It’s going to use numbers that aren’t right for development, and it doesn’t take getting the numbers too far out to make a huge difference to a feasibility. It’s the difference between a deal and not a deal. Or even worse, you think you’ve got a deal, you go ahead, only to find out it’s not a deal, and it ends up costing you a fortune.

The experiment: Oceanic Drive

Aaron: As an experiment, I thought it would be good fun to use Oceanic Drive, which is a case study we did with some students as a joint venture, and one we use all the time. I’ve got all the numbers and we know they’re 100 percent correct, because we went through that deal and everything has been ticked off. I thought it would be interesting to give the case study to the AI and see where it lands: close to our numbers or not.

Hilary: So what are we going to do first? We’re going to use ChatGPT in this case. Tell me why you’re choosing ChatGPT.

Aaron: To make it an honest go, I thought we’d use ChatGPT because the new Astra model, GPT-6, just came out. I was up at three in the morning and had it within five minutes, that’s how much of a geek I am. It is an incredible model. People are saying this looks like the start of what they call artificial general intelligence, which basically means better than any human at all tasks. I wouldn’t 100 percent agree with that yet, but they’re pretty close. It also has the best speech-to-text model, and a fantastic voice mode where you can talk back and forth. You can also set the level of thinking. I’m going to ramp it up to the second highest setting, which is probably more than most people would use anyway.

Hilary: I’m just going to backtrack for the people who are real beginners, because some of you are listening thinking, what is he even talking about? Most people are using ChatGPT or Claude. If you’re in an office you might be using Copilot, which is attached to Microsoft. Most people are using ChatGPT on their phone for simple stuff, like how do I cook this for dinner. We started on ChatGPT and went over to Claude because it was moving ahead faster and was a lot smarter. But as Aaron said, the Astra model just came out and it has gotten better and faster, so he’s moving there. We all switch between the models. So as of today, September 2026, this is the model we’re using, believing it’s probably one of the best for what we’re doing. We’ll listen to this in three years and think, what were we even talking about? We’ll probably be teleporting.

Aaron: Well, I actually use both. I use Fable, and I have a multiple set-up where they speak to each other.

Hilary: I already said he’s not allowed to talk about that, because it will freak people right out. All right, Aaron, let’s rock and roll. I want you to ask not like an experienced property developer. I want you to ask like the average person would. And remember, we’ve already done this development. We sold in 2025, so we know the answer. Aaron had to make sure the AI didn’t go and look it up on his laptop, because it went straight to that.

You can just talk to the models now

Aaron: This is the thing now. Generations ago with these models they called it prompt engineering, and it was going to be the next big thing. You had to know exactly how to talk to them. They’ve gotten that good it really doesn’t matter. You can bumble. As an example, Claude’s speech to text is woeful, about every second word is wrong, and yet the model completely understands what you’re saying. The models have gotten that good.

Aaron: There are tricks you can use to get the most out of them, but what it has really become is understanding what you’re trying to achieve. You can now set these models in motion with a goal, and all of them have it now. Anthropic calls it loops, ChatGPT calls it goals, but these models can work for over a week, no problem. So if you really don’t know what you’re trying to achieve and you set the model off for a week, it’ll get there, but you’ll get something you do not want, and you’ll have spent all that time and money. It’s even more important to know what you want. Having said that, it’s easier to talk to the models now, so I’m going to use voice for this. Let me share the screen.

Screen share: what Codex and an agent actually are

Aaron: I’ve just opened GPT, and down the bottom you can see I’ve got GPT-6 Astra set on extra high. What we’re using here is actually what they call Codex. This is not the normal chat window. This is the agent part. Most people get a little scared when we say this, because it’s called Codex, so you’d automatically think it’s all about writing code. It’s the same with Anthropic and Claude, which has the coding agent. Realistically they only called them code because that’s what people were looking for and where they saw the value in selling it, and they do write really good code. But it’s really just an agent.

Aaron: So what is an agent? A lot of people struggle with this. When you’re just talking to the model and it’s talking back, that’s a chatbot. A smart chatbot. But if you want that model to actually do things, get on your computer, check your emails, jump online, research stuff for you, that’s what an agent is. It’s the brain with scaffolding wrapped around it: software, system files, memory files. A whole bunch of files that attach to the intelligence and allow it to do useful things. You can take it to the nth degree, create personalities, generalised agents. I have tons of other agents, a whole fleet of them. But to keep it simple: it’s the model, being useful. Just because it’s called Codex doesn’t mean you only use it for coding. You can use it for exactly what I’m about to show you.

Hilary: What I call change your life.

The plan

Aaron: So here’s the model. The experiment we’re running is to see how accurate it is with a development site and a feasibility we already know, one we’ve completed and have the numbers for. I’m going to use dictation mode. If you click the right-hand button you can have a full conversation, which I love when I’m driving. OpenAI have really knocked it out of the park. It feels like a sci-fi movie, talking to your computer, and it’s got emotion in its voice and talks over the top of you like a real human would.

Aaron: I’m going to give it a very simple prompt in natural language and get it to do a feasibility. I’m going to do two things. First, I’m going to get it to do completely its own feasibility, and I’ll explain the deal to it. Then I’m going to give it the three biggest numbers in a feasibility, which make up about 80 percent. I’m going to give it the benefit of the doubt and a bit of a hand, and see what it comes up with. Then we’ll check it against the real feasibility.

Aaron: There are different types of feasibility, and this is part of the education we teach. This is one of the things the model is probably going to get wrong. It’s probably going to generalise, and I dare say it’s not even going to ask me what type of feasibility. What we want is the type of feasibility we use to look at a site to make sure it stacks up.

The prompt

Aaron: Okay, ChatGPT, I want you to give me a full feasibility please, as if we’re looking at a site to know if it stacks up. I want you to take a look at 146 Oceanic Drive in Warana on the Sunshine Coast. Do the feasibility as if you bought the land in 2023 and sold in early 2025. I know you know I’ve got a feasibility sitting on my computer, so I do not want you to cheat and look at that. Come up with your own feasibility, which is a duplex pair on that site, please.

Aaron: We had to go back in time, because otherwise it would look at today’s values and be wrong anyway. These models are trained on all the data, so it will have that pre-existing knowledge.

Hilary: And the reason we’re doing that time is that’s when it actually happened, and that’s when we have our feasibility. So it’s an apples for apples comparison. Not apples and pears, or apples and cheese.

Aaron: It’s using its spreadsheet and PDF skills. “I’ll build an independent feasibility for a duplex pair.” So helpful. These models have gotten so good. Ah, it’s stumbled across the land acquisition price in public records. That’s okay, we’ll let it run with the land. For the first run it’s got the land value, an exact value you wouldn’t normally have.

Hilary: So it’s found what we bought it for. That’s probably interesting, because we’ve given it one of the big ones people are usually unsure of.

Aaron: When we’re looking at a site, we’re usually trying to figure out what we can buy it for. But it’s got that number. Now it’s come up with the 591 square metres, it’s looking at all the planning controls, and it’ll dive into the planning scheme and all of that.

Can AI do feasibilities for you?

Aaron: While this is happening: can you get it to a point where it can do these feasibilities for you? Absolutely you can, and we do. In our business we have some Mac Studios with our own models running on them, 24 hours a day, seven days a week, doing nothing but feasibilities and looking for sites. The trick is what you teach the model. You need the development knowledge to be able to do that. You need to know what a feasibility actually looks like, how you put it together, and what those numbers are. Once you train a model to do that, you can get it to do fantastic things.

Aaron: With that, though, we’ve got a model doing something and numerous other models checking the work. A model does something, another model double-checks it. Because we’re talking money here. We don’t just go, we’ve trained it, it knows what to do. We’ve got another model checking that it’s correct and making sure there are no blips. One of the hardest things is creating a land valuation engine, which I’ve spent a lot of time on, because it is inherently difficult to figure out the value of development sites. That’s a real human thing. Valuers have to do it, and that’s why we base our feasibilities on those numbers, because we need to understand how the bank is going to look at a property. How do you teach a model to do that? You can, and it’s exactly as you say, you have other models checking its homework.

The October workshop

Hilary: We’re going to be talking AI at the three-day workshop, so a wee promo. Our workshop is on the 9th, 10th and 11th of October. This year we’re covering property development from the absolute beginner right through to the end, and we’re also covering joint ventures: creative ways of doing projects when you don’t have enough money of your own, or you’ve got another project going, or you’re tied up somewhere else. We’ve never done it before, so it’s a bit special. Aaron does a whole segment on AI, and we hang out with you for the whole three days. If you’ve been thinking about property development education, now is the time to jump in. There will always be a link below. If you want to have a chat about how you can come along, book into my diary or email admin. When you’re learning stuff that moves this fast, you want to get on board now.

Result one

Aaron: Okay, we’re back. This was the example. It did find the land price we paid, and it’s saying its profit is estimated at a $125,000 loss before tax. It’s saying $2.12 million per dwelling they would sell for, but making a loss.

Hilary: So if you were to trust it, and you saw how long it took and how thorough it was, and you just believed it, you would miss out on tons and tons of deals.

Aaron: It’s a deal we’ve done. I’ll show you the actual figures. But that’s showing a loss. Not even in the ballpark. What we’ll do is give it the next prompt with some more numbers and see where it comes out, and then we’ll open the feaso.

Hilary: Remember it had the land price because it found it. Not by sneaking on your computer.

Aaron: I think it found it through the records, probably RP Data. If you saw me scroll down, there is tons and tons of information. If you’re on YouTube you can have a look at exactly those numbers.

Round two: the three biggest numbers

Aaron: Okay, GPT. Now I’m going to give you the three biggest numbers in the feasibility: the land, the construction price and the GRV. Once again, don’t cheat and look at our feaso, but use those three numbers and redo your feasibility. You’ve already got the land, that was correct, at $1.605 million. The construction was $1,689,774, and the GRV was $4.7 million. Go ahead and do a feasibility based on those numbers, doing the rest of the feasibility with your numbers.

Aaron: While that runs, I’ll open the actual feasibility. These are the real numbers when we first looked at that site: $604,000 profit at almost 16 percent, 15.84 percent return on cost. So with GPT, if you just let it do that, it was very convincing, but it wasn’t even showing a profit, it was showing a loss. If you use those numbers, with how conservative it’s being, you would never find a deal. It’s not even close. And it’s the confidence it comes back with. It types up in front of you, it makes you believe it’s true. It gave me Excel spreadsheets, it came back with all the research, it gave me comparables, the whole bit, and yet it was still wrong. I guarantee if I pulled it apart, the basis of that feasibility would not be correct.

Result two

Aaron: Let’s see if it did any better with me giving it the three big numbers. It got 11 percent return on cost.

Hilary: A little bit better. But for a duplex we would never do a deal at 11 percent. And we wouldn’t do the deal because we wouldn’t be able to get finance.

Aaron: Correct. That goes in the bin. So even giving it the majority of the feasibility, it still isn’t where it needs to be. It’s still very significantly out. It’s obviously very conservative in this area, and I’d say it’s going to be incredibly conservative in just about every area. And it looks incredibly convincing.

Aaron: And here’s the thing. When we do our feasibilities we’re using 100 percent debt, zero percent equity, and that feasibility even included equity. That’s why I told it to look at the site like a valuer would, what we call the valuer’s methodology. It didn’t do that, because it doesn’t know that’s the way to properly look at a site.

The valuer’s methodology

Hilary: Explain the valuer’s methodology for somebody who doesn’t quite understand property development.

Aaron: When we’re looking at a site, the biggest risk we have as a developer is getting commercial finance. We need to understand how the bank is going to look at that deal. The name of the game is get the money. The person responsible for telling the bank what that development site is worth is the valuer, and they have a very specific methodology. Part of what we teach our students is how to look at a site properly. If you use all these numbers and put all this equity in, it produces less interest and makes the deal look better than what the valuer comes back and tells the bank it is. So you run the risk of getting all the way to your construction phase, getting knocked back, and the bank says no deal. We don’t want to kid ourselves. We want to look at the deal like the bank does, to make sure we get the money. The AI hasn’t done that, so that 11 percent is too high. They’ve used too much equity. If the AI had done it properly, it would have been even further out.

Hilary: And just a note, when Aaron refers to the bank he means the person lending the money. It’s commercial finance, not residential bank finance.

Aaron: Correct, and most of the time we’re not even using a bank, we’re using non-banks. The vast majority of lenders we deal with as developers are non-banks. So we say bank, but we mean the lender.

The danger of confidence

Hilary: So what should we take away? The fact that it spits it out in a graph, it does feel confident, doesn’t it?

Aaron: It’s done a ton of research, it looks great, but fundamentally it’s generically looking at the site, and that’s the problem. You still need the knowledge. If you looked at that and thought, well, it’s done a ton of research, it must know what it’s talking about, you’re never going to find a deal if you’re relying on that.

Hilary: That’s a loss of over $600,000, because that deal went ahead and the profit was $604,000. But what if it’s influenced the other way?

Aaron: That’s even worse. As the saying goes, the best deals are sometimes the ones you don’t do. If it gives you a rosy picture of a deal, you go ahead and make decisions based on that. You spend all that money getting your development approvals, then you go for your construction finance, and the bank says no deal.

The takeaway

Hilary: I think our takeaway from this podcast is that you have to use AI, but there are things you can’t trust it for. It is absolutely phenomenal, but this is a general model.

Aaron: For what we’re doing, you still need that specialist knowledge, and that should make you feel good as a property developer. It means we’re not getting replaced any time soon. You should be using these as tools. Can you, with the right knowledge, get them to the point where they can create incredible feasibilities? Yes, you can. But it’s your knowledge going in there to tune the models. You still need to know what you’re doing.

Hilary: There are some really good basic questions where you can use AI to get started: populations, towns, real estate. There are so many things you can use it for. But the big stuff you have to understand.

Aaron: Absolutely. Town planning is another one. Planning schemes. No longer do you have to sit there reading through town plans and codes, which can really do your head in. For things like that, AI is fantastic.

Hilary: So hopefully we’ve inspired people to jump on board and have a crack, even if it just means downloading ChatGPT onto your phone and starting to play with it. Talk to it. If you’ve not talked to GPT lately, you’ll be blown away by how good the conversations are. It doesn’t replace people, and it doesn’t replace a lot of things, but it is very useful. It’s like the car, like electricity, like the internet. It is coming, it is here, so get on board. But a great demonstration today, Aaron, showing people you still have to have development knowledge. If property development is something you want to get into, we are the people to learn from, because we are incorporating AI into our learning, because we use it ourselves, and we push people to use it, because otherwise you’re going to get left behind.

Aaron: And to emphasise your point about not being scared of AI: have a go, and give it the hardest possible thing you can think of. Push the boundaries. But as we’ve demonstrated today, there is absolutely still the requirement for specialised knowledge. If you can take your specialised knowledge and use AI, it’s like real estate development on steroids.

Hilary: Aaron, thanks so much, as usual. I love that you love talking about AI so much, and I’m so grateful you’ve taught me so much. Usually you leave the office and I have to start asking ChatGPT, what did Aaron mean by that? We hope you got something out of it. As I said, our workshop is coming up on the 9th, 10th and 11th of October. It is going to be a phenomenal event, incorporating property development, joint ventures and AI. You will walk away a new person. If you want to know more, reach out, there’s always a link below. We’ll catch you next week. Bye for now.