5,000 Hours of AI in Real Estate: 3 Things the Best Guide I've Read Actually Gets Right
The AI conversation in real estate is loud. Most of it isn't useful.
Welcome to a new issue of the Unlocking Real Estate Value newsletter. Each week I will provide you with exclusive advice and professional insights to help you realise long-term value through real estate development.
This week, Francis Huang, ex-real estate private equity, seven years of AI and asset pricing research at Harvard and MIT, published something rare. A free guide to AI in CRE built on 5,000+ hours of real deployment at institutional funds managing real capital on real deals.
Not a vendor pitch. Not a think piece. Mechanics.
I read it.
Here’s my filter, and what it means if you work in real estate development.
Lesson 1: The model doesn’t understand your document. It generates text that looks like it does.
This is the most important thing to grasp before you use AI on anything that matters.
Large language models generate output token by token, based on statistical patterns. When you upload a planning consent, a construction contract, or a feasibility study, the model isn’t reading it the way a lawyer or analyst would. It’s producing text that resembles a correct answer. Not the same thing.
Two practical consequences follow. First, context windows: models have hard limits on how much text they can process at once. Feed in an 80-page document and ask for all key data, and the model may silently drop sections without warning. Deadlines, floor areas, suspensive conditions. Gone, with no error message. Second, hallucinations: the model may invent a figure, smooth over an inconsistency, or paraphrase away a restrictive clause. The output looks identical to a correct one. Confidence tells you nothing about accuracy.
The developer who doesn’t know this treats AI output as a draft to refine. The developer who does builds a protocol around it: explicit rules in the prompt, source mapping, placeholders for every undocumented field.
Protocol replaces trust. This matters everywhere. It matters more in markets like Italy, where input documents are often scanned PDFs in bureaucratic language with ambiguous clause structures. Higher risk, smaller margin for error.
Lesson 2: Where AI works in development, and where it quietly fails you.
Most AI coverage tells you what’s possible. Huang tells you what’s reliable. That’s a different conversation.
Here’s the honest map, filtered for development work. Offer memoranda and investor communications work well for extraction, synthesis, and reformatting, with anti-hallucination protocols in place. Lease abstraction works on structured documents; fails on ambiguous clauses or non-standard language. Market research works for aggregation; fails on real-time or hyper-local data. Don’t use it to generate comparables you can’t verify independently. Financial modelling is where the guide marks the most significant shift, covered in Lesson 3.
Without this map, you deploy AI where it fails. You chase use cases that sound impressive but produce unreliable output. With it, you build a workflow around AI’s actual reliability curve. You stop experimenting and start implementing.
The question shifts too. It stops being “can it do this?” and becomes “is this a use case where I can trust the output enough to act on it?”
Lesson 3: Something shifted in late 2025. The analyst’s role changed.
Huang documents a qualitative change in the second half of 2025: purpose-built real estate tools can now produce institutional-grade Excel models from natural language input. Full DCF structure. LP/GP waterfall with IRR hurdles. Articulated debt stack. In minutes.
This doesn’t eliminate the analyst. It eliminates the part of the work that consumed time without adding judgment: building the skeleton.
For a development pre-feasibility, that’s significant. Cost planning, revenue simulation, scenario analysis. Assembly work that took days now takes hours. The analyst stops building the model and starts reasoning about it. Earlier in the process. Across more opportunities.
For a development team without a dedicated analyst, which describes most of the market in Italy and elsewhere, the implication is direct. You don’t need to hire before you can model. You need to verify what the model produces.
The question worth sitting with
Huang closes with this: the professionals who navigate AI well won’t be the ones who adopt it fastest. They’ll be the ones who understand it well enough to know when to trust it, when to verify it, and when to push back.
Not adoption speed. Calibration.
Key takeaways:
AI generates text that resembles correct answers. It doesn’t understand your documents
Context windows and hallucinations are operational risks. Build protocols, not habits
Deploy AI where its reliability is proven; approach other use cases with structured scepticism
The late-2025 shift on financial modelling materially compresses pre-feasibility work
The guide is free: apers.app/library/guide-to-ai-in-cre
That’s all for today. See you next week.
— Carlo
Founder and Managing Director Benigni
Sanity Check Your Project
Before you commit another million to your development, stress-test it with someone who’s delivered 2M sqm and €11B+ GDV across Europe.
Here’s what I can pressure-test for you:
Is your market thesis still valid? Office, hospitality and residential markets are shifting fast. What worked 18 months ago might not work today.
Will your budget survive the current cost environment? Construction costs and programme risk need a reality check before you commit.
Does your ESG strategy attract or repel capital? Tenants and investors have moved past good intentions. They want metrics.
What risks are you not pricing in? Regulatory shifts, planning delays, cost inflation. The ones that kill projects are the ones you didn’t model.
Book a free 15-minute strategy call. One conversation. No commitment. Just clarity on your next move.
Refer a Colleague
PS…If you’re enjoying this newsletter, please consider referring this edition to a friend or a colleague. Sharing valuable insights helps everyone make better investment decisions.
This post is sponsored by Benigni a specialist development manager working with international investors to realise long-term value through optimised development strategies. To learn more click this link to our website.
Subscribe to the Newsletter
A newsletter by Carlo Benigni providing, in less than 4 minutes, exclusive advice, strategies, and resources to help unlock hidden real estate value.
Subscribe for free to receive new posts and support my work.
Interesting. This is consistent with our findings so far. The narrower and better defined the question, the more we trust the answer. And always ask for the source paragraph so you can quickly eyeball the underlying text.