The past few years has seen the proliferation of tools that seek to automate aspects of real estate development. From deal sourcing to underwriting to cost estimation and procurement, there’s now an AI-enabled app out there for almost every step of the development journey.
To date there has been no machine-led real estate development firm. But it’s at least theoretically possible to stitch these apps together to automate the process end-to-end. So how far away are we from this theory becoming reality?
Today’s letter will attempt to answer that with a thought experiment: if I were to build a real estate development firm with no humans, how would I go about that? What apps would I use and how would I design the process?
Today we’ll design and launch Asimov Partners, a hypothetical real estate development firm that builds ground-up multifamily but has zero human employees doing any actual work. As we tackle this, we’ll touch on some of the technology companies pushing the boundaries in automation as well as what this might all mean for real estate development as a profession in 3, 5, and 10 years.
Sketching it All Out
Building Asimov Partners’ technology stack requires breaking the real estate development process down into its component parts. Deal sourcing, underwriting, zoning analysis, cost estimation, et cetera—each has its own ecosystem of AI-powered apps and services. Lots of them overlap, so we’ll pick a representative tool or two from each category.
Of course, these technologies don’t do things without being told to do them. Remember, Asimov Partners doesn’t have any humans to tell the robots what to do. So we’ll need a “pilot” to tie our technologies together and make them work. For that, we’ll use GPT 4o with lots of custom instructions. It will be able to manipulate basic files and give instructions to our various tools, and it’ll play a more active role as we get deeper into the development process.
For today’s example, Asimov Partners will focus on ground-up multifamily for two reasons. One, apartment development is relatively complex in comparison to (say) light industrial or retail. Two, I’m a little closer to it so can speak more directly to how these technologies would be applied.
Deal Sourcing
To get the process kicked off, Asimov Partners needs deals to evaluate. There are two possible strategies to get deals in the top of the funnel:
- Scrape on-market deals from LoopNet and run them through the process we’ll describe below, treating list price as an input and evaluating deals from there.
- Use a zoning analysis tool like CityBldr to find off-market sites with development potential. New tools like Permit Portal (YC F24) promise to make this process even easier by finding high-potential sites with natural language input. (If we go this route, we’ll need to automate the generation of offers for each viable site and use a direct mail automation tool like Postalytics to send offers out to owners. All very doable.)
Regardless of the strategy Asimov Partners chooses, the output of this step is the same: a list of potentially interesting addresses with development potential.
Zoning Analysis & Test Fit
As a next step, we’ll need to validate what we can actually build on each site. Depending on the deal sourcing strategy we chose in the prior step, we may already have insight here, but we’ll assume for the moment we need to start from scratch.
Fortunately, there are a number of tools that could help us out here. CityBldr can provide us with zoning information including buildable units and parking requirements as well as environmental indicators. Gridics can also produce a detailed zoning analysis report, although they charge per report ($1,499) so perhaps we’ll layer that in once we’ve honed in on a smaller number of viable development sites.
But we’ll likely go with Zoneomics for this step for one reason: they have an API, so we can automate getting addresses in and development potential out.
A proper underwriting, however, will require more than just a high-level unit count and a massing. To get the next level of detail we’ll spin up TestFit’s Site Solver, which can translate our zoning data into a detailed site plan, unit mix, and even quantity takeoffs. Generating material quantity takeoffs is critical, as those are a key input to the construction estimation we’ll do later on. TestFit is a key part of our site plan automation, as it can run through a lot of options quickly.
From TestFit we’ll get a detailed unit mix, floor plans, and material takeoffs to feed into the next steps of the analysis.
Underwriting
The financial modeling part of this exercise is relatively straightforward. With our detailed unit mix in hand, we’ll use Archer’s “bring your own model” feature to automate the creation of a property pro forma and financial model. We’ll also use Archer’s data as well as HelloData’s market surveys generate comps and come up with rent assumptions. (We’ll use HelloData’s API to pull market data into Archer.)
The end result of this will be the revenue and opex side of a property pro forma—interesting, but we’ll need the cost side as well to make an investment decision.
Cost Analysis
While it’d be possible to use GPT to reach out to real construction firms and get actual estimates of construction cost, Asimov Partners is planning to run thousands of sites through this process—so getting any humans involved at this stage is unrealistic.
Instead, we’ll spin up Ediphi to get a high-level estimate of construction costs. Ediphi has regional costs databases and will allow us to turn our plans for each site into high-level construction estimates. Ediphi can receive takeoff data from TestFit, so it’ll know our site and unit plans as well as the specific quantities of materials we’ll need for core and shell.
For estimating and procuring all the finishes and fixtures, we’ll use GoSource’s materials estimator. They have a handy tool that generates a list of actual SKUs and prices from any interior image, including interior images generated by AI. So we’ll use GPT to create images of the interior aesthetic we’re going for; feeding that into GoSource will create a list of SKUs with pricing. Combining that with the takeoffs we received from TestFit will give us quantities of each SKU.
With any luck, we now have detailed pricing estimates of core and shell (including MEP) from Ediphi and fixtures and finishes—specific SKUs that meet our design standards—from GoSource. To wrap it up, we’ll combine these construction cost estimates with our pro forma model in Archer to produce a complete project financial model.
A Stopping Point?
At this point, we have a list of development sites with massings, test fits with unit counts, material takeoffs, and pro forma P&Ls showing projected cost, revenue, and NOI. From there, it’s relatively trivial to calculate projected unlevered yield on cost and IRR for each project. We can then apply some basic heuristics to narrow down our list—say, eliminate any project showing a <7% UYOC or <18% IRR. This simple filter would leave us with a list of “qualified” sites.
For most developers, this would be a reasonable stopping point. Our analysts can spend their time evaluating pre-qualified development sites, poking holes in our tools’ analysis or finding opportunities the software missed. On its own, this stopping point would be a revelation, offering significant cost savings and allowing a developer to evaluate far more sites and find opportunities they might otherwise have missed.
Of course, this is a thought experiment. So there’s no reason to stop here.
We’re gonna keep going and build this thing.
Welcoming back GPT and Building the Thing
At this point, we need to replicate the obvious thing any human analyst would do: go back and check the computers’ work, especially when the computer produced outlandish results—30%+ IRRs, 10%+ UYOCs, or atypical average unit sizes, for instance.
I’m not aware of any good specialized tool for this since it solves a problem that doesn’t yet exist. So we’re going back to GPT 4o—our generalist tool—and ask it to put on its real estate finance analyst hat and find problems. (The more custom instructions we give it here, the better, and training it on ‘typical’ development models and financial returns should help it find outliers.)
This will necessarily be an imperfect process. Tightening the aperture against false negatives (bad deals that get through) will inevitably lead to more false positives (good deals that get screened out), making the end result less interesting. The human creators of Asimov Partners will need to pick their level of risk tolerance and instruct GPT accordingly.
Buying the Site
Once we’ve identified our target site, we’ll need to put in an offer on it. Following a set of instructions to negotiate a deal is fairly trivial work for an LLM; interfacing with a human broker (or seller) is a bit more complicated.
Automating the back-and-forth via email is pretty straightforward and bog standard for any LLM. Dodging requests for phone calls might be a little weird, but I trust GPT to come up with some solid-sounding excuses.
And while it’s a little tougher to dodge an in-person presence at the closing table, there’s no requirement that the buyer show up in the flesh as long as someone can sign the documents. This is, of course, complicated given that people and business entities cannot give power of attorney to an LLM.
I’m going to hack this by assigning PoA to an entity that is controlled by an AI agent with a human-sounding name, which “signs” the document as the entity’s managing member. Would this hold up in court if it were challenged? Probably not! Would our seller notice or care? Also probably not! (Asimov’s lender may not be so sanguine.)
Hacks aside, this is a genuine gap in our thought experiment. Without a solid way of giving the LLM signing authority, a human would likely have to step in to sign legally-binding documents. Of course, we might just fudge a bit and hire a third-party human lawyer with PoA up-front and strict instructions to sign whatever the LLM decides.
Raising Capital
Assuming Asimov Partners is not sitting on a pile of cash, we’ll need to raise both equity and debt to finance these projects. We’ll use StoryDoc’s AI pitch deck creator to automate the generation of marketing materials for our deal using the key information from our financial model, site plan, and unit mix. And we’ll use mnml.ai to translate the architectural information we already have into beautiful renderings for our marketing materials.
Asimov Partners will use crowdfunding platforms to get our deal in front of investors. Fundrise actually has an API, although it’s not clear if it can be used to add deals to the platform or simply pull data from already-listed deals.
At this point we have a fully AI-generated real estate project in front of real investors. We think they’ll eat it up.
Permitting and Construction
Once we’ve capitalized our deal, we’ll pull permits using a software-based service like PermitFlow, which bills itself as “TurboTax for construction permitting.” Unfortunately, PermitFlow doesn’t appear to have an API, so it’s not clear we can fully automate this step.
Unfortunately, there aren’t really many good options for automating the management of an actual construction process. While software like ProCore’s bid management tools can make the process easier, they still require some human involvement.
So we’ll have to lean heavily on GPT to find and communicate with construction firms as well as run the bidding process. Once we’ve selected a GC and construction has started, we can try out some bleeding edge technology like Percepto to monitor construction progress.
Leasing, Management, and Sale
There’s a lot that we could dig into when it comes to automating multifamily leasing and property management—enough, in fact, for an entire letter. So we’re not going to cover that here; instead, Asimov Partners will outsource property management to a third party. (After all, our high-throughput site analysis works best if we’re not constrained by the markets in which we have first-party property management teams.)
Once our site is fully leased, we’ll instruct GPT to hire an investment sales broker to shop it. Naturally, offers will be evaluated and negotiated by our LLM. With any luck, we’ll be able to return capital to our backers over on Fundrise.
So how far away are we?
While this was a fun thought experiment, there are clearly gaps in the automation process, particularly after we picked a site and launched into the actual financing and construction work. There aren’t really any good ways to automate dealing with a lender, for instance, and construction bidding and monitoring tools all still assume there’s a human piloting the system. While there are some hacky ways around that, they’re not going to produce great results and will require a lot of maintenance.
But for most developers, those gaps matter far less than the quality of end product. And stringing multiple automated systems together—even if possible—is probably going to introduce errors and come up with some wacky results.
To get an expert’s take, I spoke with Murat Melek, Director of Artificial Intelligence at Suffolk Construction. Melek is charged with researching and implementing AI at one of the largest construction contractors in the country, so he sees a lot.
“We’re nearly there with automation, but it’s not yet sufficient,” said Melek. “Current tools will help you come up with a layout. Where the apartments are, where the doors are. But then you need to do mechanical load calculations. How much air do I need to pump into the space? How many diffusers will I need? What sizes, and how are they connected to units with ductwork? Same with plumbing and sprinklers, for instance.”
Much of the work of an architect, after all, is coordinating between various trades, something that technology hasn’t yet cracked. “We also need to calculate structural depth and how it’s all going to fit together,” notes Melek. “TestFit might say a 15’ floor is okay, but is it? There could be huge implications if you go from 15 to 15’6”. You need to go to the structural engineer, to the plumber, et cetera to coordinate and align on changes.”
Still, Melek sees potential in many recent advances, particularly the potential for streamlining communication between trades and automating rote, time-consuming architecture and engineering tasks.
As we see with Asimov Partners, the usefulness of automation at any given task depends on (a) the volume at which it needs to be performed and (b) the required fidelity of the output. For this reason, automation is most useful as we go further up the funnel. The first few steps we laid out here—tying together site identification, zoning analysis, test fit, and underwriting—is by far the most useful. Most developers would be happy to receive a hit list of promising sites they might not otherwise have seen and do the rest the old fashioned way.
And it doesn’t seem like we’re far at all from those first steps being a reality. Tools like TestFit, Tailorbird, Cedar, GoSource, and and Ediphi have made tremendous progress in recent years, so much so that’s they now have overlapping features and could clearly be strung together. They’re not, however, built with that use case in mind—yet—so actually getting the data from one platform to another would likely require some finagling. But that is, in a way, the easy part.
It’s hard to overstate the impact that this could have on what it means to be a real estate developer. For one, it fundamentally changes the job of an analyst, replacing the rote work of underwriting with more time on validation and analysis. With high-throughput deal analysis available to anyone, best development opportunities will be found by those firms able to craft the best screening algorithms, ensuring that the highest-potential deals get in front of a human analyst.
From Melek’s perspective, AI also has the potential to remove a major human bottleneck. “Right now it’s hard to find engineers. There aren’t enough of them.” By automating rote work and allowing architects and engineers to focus on the highest-leverage tasks, AI has the potential to mitigate this skills gap.
“If you’re not designing something very complex—like the next high-rise in a seismic zone or an art museum, for example—for most typical office and residential buildings, most of of the design can be automated,” said Melek.
While quant real estate firms like Stablewood and Two Sigma make heavy use of data and automation to inform real estate investment decisions, they’re mostly doing so at a neighborhood or census block level—no one, as far as I’m aware, is automating the actual development process. For now, Asimov Partners is only a dream—but perhaps not for much longer.
This article was originally published in Thesis Driven and is republished here with permission.
