Table of Contents
Over two recent episodes of Real Estate Meets AI, we spoke with three researchers who spend their time studying real estate from outside the day-to-day work of building, financing, and operating it.
Arpit Gupta at NYU Stern studies real estate markets, urban economics, and housing regulation. James Robert Scott at MIT tracks more than 8,000 proptech companies and studies how technology moves through the industry. Sam Chandan, also at NYU Stern, works at the intersection of real estate, finance, and AI.
We went into those conversations asking where AI is actually making a difference in real estate. But as we listened back, a different pattern became hard to ignore: again and again, the conversation came back to why real estate has been so difficult to change in the first place.
That question matters because the industry is, by most measures, behind. BCG found that only 25% of real estate firms qualify as AI leaders, compared with 40% across industries. In 2026, real estate firms are also investing roughly half the cross-industry average in AI.
It’s easy to chalk that up to an industry that is simply slow to change. Our conversations suggested a more complicated explanation. Real estate has a data problem, an economics problem, and an adoption problem rooted in the very expertise that makes the industry work. Taken together, those three constraints help explain why AI has moved faster elsewhere, and what will have to change for real estate to catch up.
Real estate wasn’t built to produce machine-readable knowledge
Gupta’s work offers a useful case study in the problem. His team recently built AI tools to analyze zoning regulations across U.S. municipalities, where even seemingly straightforward questions become difficult to answer consistently at scale.
They set out to answer roughly 20 structured questions about those regulations: minimum lot sizes, multifamily permissions, setback requirements, density restrictions, public hearing requirements, and more. These are basic questions for anyone evaluating a site. Getting reliable answers at scale is anything but basic.
Zoning codes are scattered across thousands of jurisdictions. They are formatted differently, written differently, stored differently, and often buried in dense legal language. There’s no universal database and no consistent schema. Each jurisdiction effectively presents its own information problem.
Anyone who has dealt with the cost of a zoning surprise on a live project already knows the consequence of that fragmentation.
When Gupta’s team initially fed those documents into a general-purpose AI tool, he estimated accuracy at roughly 60%. That’s not good enough when the answer could influence whether capital gets committed to a site. To get above 90% accuracy, with traceable citations back to the source text, the team had to build a custom retrieval pipeline that first identified the relevant section of each document before querying it.
The distance between 60% and 90% is the data problem, quantified.
Zoning is only one example. Real estate work moves constantly between PDFs, drawings, spreadsheets, images, financial models, legal documents, emails, market data, and information that may never have been digitized at all. Scott describes an industry where data remains siloed across functions and stakeholders, and where the information itself is often too local and nuanced to standardize easily.
That matters because AI adoption has moved fastest in environments where information was already relatively structured. Software had enormous repositories of machine-readable code. Finance had decades of digitized transaction, pricing, and market data. Those industries didn’t move faster simply because they were more willing to innovate. They started from information environments that made early AI tools more useful.
Real estate didn’t have the same advantage.
Chandan adds a human layer to the problem. A meaningful portion of the industry’s workforce built their careers before data-native workflows became standard, so the challenge is not just converting information into a format AI can use. It is redesigning processes that were never built around structured information in the first place.
Real estate’s data problem is not a temporary backlog of messy files waiting to be cleaned. The industry’s knowledge itself is fragmented, multimodal, local, and distributed across organizations and people. AI has to operate within that reality.
The firms with the most to gain from efficiency often have the least room to experiment
The second constraint is economic.
Scott points to construction as the clearest example. Construction often operates on margins of roughly 2.5% to 4.5%. At those margins, a failed technology investment is not an inconsequential R&D expense. It can materially affect the performance of the business.
The result is an uncomfortable paradox: the firms facing the greatest pressure to become more efficient may also have the least capacity to experiment their way toward that efficiency.
The pressure is getting worse. Scott cites a 12.5% increase in construction costs since January, layered on top of already elevated baselines. Tariffs and material costs are compounding the problem.
It would be reasonable to assume that rising costs should accelerate technology adoption. In practice, they can do the opposite. An organization operating on thin margins has less tolerance for a tool that requires months of testing, workflow redesign, training, and iteration before it produces a reliable return.
Scott also describes the pressure from the organizational side. Outside of a few unusually strong sectors like data centers, he says, it has been a rough three years for the industry. Firms have fires to put out on multiple fronts, and leadership teams already have enough on their plate before anyone asks them to rethink how their organizations work.
A lot of executives, Scott observes, want three bullet points and an AI strategy. But identifying where AI can actually create value requires the opposite: self-reflection, workflow analysis, an honest look at where the bottlenecks really are, and enough organizational capacity to change the answer. Those activities compete for the same time and capital already being consumed by the core business.
Chandan frames the challenge from the talent side. As AI-equipped analysts become dramatically more productive, firms are discovering that the same volume of work can be handled by fewer people. In a finite deal environment, the result is not more output from more analysts. It is the same output from fewer. That dynamic makes the return on AI real, but it also introduces new questions about team structure, hiring, and how firms invest in their people during a period of rapid change.
Gupta’s comparison with finance is useful here. Many financial firms entered the current AI cycle with established R&D cultures, meaningful technology budgets, technical talent, and data infrastructure already in place. Most real estate organizations didn’t. They are being asked to experiment with a new class of technology without the economic cushion or technical foundation that made experimentation easier elsewhere.
In real estate, expertise creates its own switching costs
The third constraint may be the most misunderstood.
It’s easy to characterize real estate’s slow adoption of technology as conservatism, but that misses why established workflows have survived. Real estate expertise is unusually dependent on experience.
Scott describes development as deeply local. Successful developers know their cities. They know the planning department, the lenders, the contractors, the political dynamics, the neighborhoods, and the people who can derail or advance a project. Many of those relationships have been built across multiple market cycles.
That knowledge is not a workaround for missing technology. It is a competitive advantage, and that creates a significant switching cost. A new AI workflow is not competing against nothing. It is competing against systems, relationships, and instincts that have produced real economic value for decades.
Scott argues that this is part of what outsiders often underestimate about the industry. Real estate can look simple from a distance: acquire land, finance a project, build it, lease it, operate it. In practice, every one of those stages sits inside overlapping physical, regulatory, financial, political, and human systems. Playbooks imported directly from software or finance rarely survive contact with that complexity.
Chandan sees another side of the same issue. People across the industry have very different incentives to embrace AI. Leaders, academics, and technologists may see productivity gains and new capabilities. Someone whose economic value has been built around a particular skill or process may see something very different.
That divide has real implications for adoption. A tool can work perfectly on paper and still fail if the people expected to use it do not trust it, understand it, or see how they fit into the workflow it creates.
There is also a more fundamental tension at play. A bad line of code can be rolled back. A bad building cannot. Projects consume enormous amounts of capital, unfold over years, involve dozens of stakeholders, and produce physical consequences that are expensive to reverse. The industry’s resistance to “move fast and break things” is, in many contexts, entirely rational.
AI, however, has largely developed inside a culture of iteration: test, fail, adjust, repeat. Real estate needs that experimentation but operates in an environment where experiments can be unusually expensive.
These problems reinforce each other
The data, economics, and culture problems are not independent. Messy data makes early AI tools less reliable. Less reliable tools require more experimentation. Thin margins make that experimentation harder to fund. And disappointing early results reinforce the instincts of experienced practitioners who already have functioning ways of doing the work.
That is why the gap between AI awareness and AI adoption can persist even when virtually everyone agrees that the technology will matter. It also helps explain why comparisons with other industries can be misleading.
Real estate is not simply several years behind software on the same adoption curve. It is trying to adopt the technology under a different set of conditions. BCG’s own analysis points to structural challenges, including fragmented data landscapes and historically asset-centric operating models, as contributors to the industry’s slower adoption.
That means real estate probably won’t become AI-native by copying the way software became AI-native. It will need an adoption model built around the realities of the industry itself.
The clock is still running
None of this means the industry is standing still. Gupta’s zoning research shows that tools can already solve problems that would have been extraordinarily difficult to tackle at national scale only a few years ago. Chandan sees an industry more open to experimentation than it has historically been. Scott is tracking thousands of companies trying to modernize nearly every part of the real estate lifecycle.
The capabilities are improving, and that matters because the barriers outlined here are not fixed. Fragmented data can be made more usable. Workflows can be redesigned around AI rather than simply layering new tools onto old processes. And adoption can happen in ways that preserve the judgment, relationships, and local expertise that make real estate work in the first place.
Chandan compares the current wave of AI investment and adoption to the railroads, not because the technologies are equivalent, but because of the scale of economic change he believes is underway. The more important distinction may be speed. Past technological revolutions gave industries years or decades to reorganize around them. Today’s AI capabilities are changing on a much shorter cycle, which changes what being “slow” means.
When we asked the three researchers where real estate sits on the AI adoption curve, none described an industry close to maturity. The exact score mattered less than the pattern: awareness is widespread, experimentation is uneven, and meaningful integration remains limited.
The answer is not simply to tell real estate firms to use more AI. It is to build AI around the realities of real estate.
That means systems capable of working with fragmented and multimodal data rather than assuming clean inputs. It means business cases that can survive without enormous R&D budgets. And it means workflows that amplify the judgment, relationships, and local knowledge practitioners already have rather than trying to replace them.
Those are difficult constraints, but they are also an opportunity. Industries do not get transformed by adopting someone else’s playbook. They change when the technology is built to be useful on their terms.
Real estate has been slower to adopt AI for real reasons. The next phase will depend on whether the industry can turn those constraints into the blueprint for how AI gets built, adopted, and trusted here.
This piece draws from Season 2, Episodes 4 and 5 of Real Estate Meets AI. Listen to the full conversations with Arpit Gupta and James Robert Scott and with Sam Chandan wherever you get your podcasts.