I've watched a lot of people try to use Claude/ChatGPT on real estate models. The pattern is almost always the same: it works long enough to look impressive, then quietly falls apart somewhere between tab six and tab twelve.
This isn't a criticism of the AI labs. OpenAI and Anthropic have built genuinely remarkable systems. But real estate financial modeling is, I've come to believe, a genuine edge case for them: one that doesn't show up clearly in their benchmarks and doesn't fit neatly into how their tools were designed to work.
Here's why.
A real estate model is not (just) a spreadsheet
When most people hear "AI for spreadsheets," they imagine something like a corporate FP&A model, a budget template, or a hedge fund's return attribution table. These are essentially tabular data structures — relatively flat, column-labeled, logically sequential. You can read them like a document. A CSV file captures most of what matters. The structure is clean. Usually the analysis is a few lines of Python codes after you imported pandas library.
A commercial real estate Excel model is something else entirely, to a point it feels like a computational system: built block by block, tab by tab, over months or years, by people who each left their own fingerprints on it. It has assumptions scattered across a dedicated inputs tab, formula chains that snake across eight intermediate tabs before surfacing in a summary, waterfall structures that reference partnership terms defined somewhere in the middle of the file, and sensitivity tables that pull from all of it simultaneously.
The navigation is inherently visual. An experienced analyst reads a real estate model the way a contractor reads a set of architectural drawings, not linearly, but spatially. They know where things live. They know which blocks feed which. They know that the named range LTV_Senior in the debt tab means something slightly different from the field labeled the same way in the assumptions tab. They hold the whole structure in their head.
General AI has no equivalent of this spatial comprehension. It reads cells. It processes ranges. It sees the words but misses the architecture.
The information problem is worse than you think
Even if you could teach an AI to navigate the visual structure of a real estate model, you'd immediately run into the second problem: the sheer volume of information required to build one correctly.
A typical acquisition involves a data room — lease documents, operating statements, rent rolls, an Argus export, maybe a prior appraisal and a few broker OMs. Across a moderately complex deal, that's easily hundreds of pages of source material that need to be read, extracted, reconciled, and mapped to the right cells in the right model structure. And it all has to happen coherently — not in isolation, but as an integrated financial picture where every input is connected to every output.
The context windows of today's best large language models, impressive as they are, are not designed for this. You can push a long context into GPT-5.4 or Claude 4.6 and get a reasonable answer about a single document. But when the task is to read a 40-page lease, cross-reference it against a rent roll, reconcile both against a prior year operating statement, and then populate a model that has its own internal logic across thirty tabs — the window fills before the model is half-assembled. And when a language model runs out of context, it doesn't tell you. It just starts to drift.
This isn't a complaint about any particular product. It's a structural observation about what the problem requires. The information density of a real estate deal exceeds what current LLMs can hold in focus simultaneously. Full stop.
Why this matters
The companies building general AI have enormous surface area to cover. Real estate financial modeling is a niche within a niche — a small slice of the total addressable use case for a foundation model. It makes complete sense that it hasn't been solved by the platforms building for everyone.
But for the people who work in this industry, it is not a niche. It is the core artifact through which capital decisions get made. The financial model is where risk gets quantified, where returns get projected, where a deal gets a yes or a no.
That gap — between how important the model is to practitioners and how poorly equipped general AI is to handle it — is the gap I keep thinking about.
The problem isn't that AI isn't good enough. It's that real estate modeling is a genuinely different computational environment than the ones general AI was designed for. Recognizing that is the starting point for building something that actually works.
