How David Bellairian Uses AI to Underwrite a Rental Property (and Where He Stops)
I’m David Bellairian. I run Glencore Associates, a private investment and asset management firm in Los Angeles, and I founded AIDiscover, an AI marketing agency. I use AI models every day, and I have written before about how I use them for stocks. This essay is about the other side of the portfolio: how I use AI when I am looking at a rental property, and, just as important, where I stop letting it help.
The rule before the method
The rule is simple. The model never makes the decision, and it never supplies a number I have not verified. It is a research assistant with an unlimited attention span and no judgment. If you give it the role of analyst, it will happily play one, and it will be confidently wrong in exactly the places where being wrong costs the most. If you give it the role of junior associate, it will save you ten hours a week.
Everything below follows from that.
Step one: get the facts out of the documents
A property comes with a stack of paper: the listing, the rent roll, the leases, the seller’s expense statement, the inspection report, the title preliminary, sometimes a permit history. The first thing I do is have the model read all of it and build a single structured summary: units, current rents, lease terms and expirations, stated expenses by category, any deferred maintenance the inspector flagged, anything in title that looks unusual.
This is where the model is at its best. It is reading comprehension at scale. It does not get bored on page forty of an inspection report, and it will catch that the rent roll says one thing and the lease says another. I ask it specifically to flag inconsistencies between documents rather than smooth them over, because the inconsistencies are usually where the real story is.
What I do not do is accept any of those numbers as true. Every figure that matters gets checked against the source document by me. The summary tells me where to look. It does not tell me what is there.
Step two: stress the seller’s story
Seller-provided expenses are a sales document. They are almost always too low. So the second thing I ask the model to do is act as a skeptic: given this building’s age, size, and location, what expense categories look light, what is missing entirely, and what would a conservative operator budget instead?
The model is good at producing the checklist of things people forget: reserves for capital items, vacancy and collection loss, management even if you self-manage, insurance that has repriced sharply in California, utilities the seller may be passing through today but may not be able to after a tenant turns over. It is not good at the actual numbers, because it does not know what my insurance broker quoted last month or what a roof costs in this zip code this year. The categories come from the model. The figures come from my own records and from people I call.
Step three: the regulatory layer
In Los Angeles this step decides whether a deal exists at all. Rent stabilization, just-cause eviction rules, relocation assistance, local registration requirements, and the state’s Costa-Hawkins framework all change what a building can earn and how quickly. I covered how that shapes what I actually buy in an earlier essay.
Here I use the model carefully. It is useful for orientation: which rules might apply given the building’s age and the city it sits in, and what questions to ask. It is not a source of truth. Rules change, local ordinances layer on top of state law, and a model’s answer can be a year stale or simply invented. So I have it draft the list of questions, and then I answer the questions from the city’s own published ordinance text, the state statute, and, where money is on the line, a licensed professional. I hold a law degree, which makes reading the statute easier for me than it is for most operators, but the degree is not a substitute for confirming the current version of the rule.
Step four: the physical risk I learned the hard way
I once bought a Glendale property without testing for lead and asbestos, and the bill taught me something that is now permanently in my process. For any building older than roughly 1980, I have the model build an environmental due-diligence checklist from the inspection report and the building’s age: what should be tested, what the inspector did not look at, which disclosures the seller is obligated to make.
Again: the model builds the checklist. A licensed inspector does the testing. The model’s job is to make sure I do not forget to order the test. That is a small job. It is also the one that would have saved me a great deal of money.
Step five: the underwriting itself
Only now do the numbers go into a model of the deal. I build it myself, in a spreadsheet, with the verified rents, my own expense assumptions, a defined recovery window, and conservative exit assumptions. The AI’s role at this stage is to review, not to build. I give it the completed underwriting and ask it to find the assumptions that are doing the most work, to tell me what has to be true for the deal to perform, and to argue the other side.
That last part is the most valuable use of the model in the whole process. It is tireless at devil’s advocacy. It will tell me the rent growth assumption is optimistic, that the exit cap rate is thin, that I have not budgeted for the vacancy a renovation creates. Sometimes it is wrong. Often it has found the thing I was quietly hoping no one would ask about.
Where I stop
I do not let the model estimate market rents. I pull comparable listings and recent leases myself. I do not let it estimate renovation costs. I get bids. I do not let it read a lease and tell me what the lease means legally. I read the lease, and if it matters, counsel reads it. And I never let it tell me whether to buy. By the time the question is “should I do this deal,” the model has done its work. The decision is mine, with my capital, and I want it to feel that way.
If you take one idea from this: use AI to make sure you asked every question. Do not use it to answer the ones that cost money if the answer is wrong.
This essay describes how one private firm evaluates property for its own account. It is not investment, legal, or tax advice. I am not an attorney, and nothing here is an offer or solicitation. Glencore Associates does not manage capital for outside investors.
— David Bellairian, Founder & CEO, Glencore Associates | Founder, AIDiscover | davidbellairian.com
More from David Bellairian: How David Bellairian Sizes Gold in a Portfolio (Substack) · David Bellairian on AI Stock Trading (Medium) · Business Book Recommendations (Glencore Associates) · Essays on Investing, AI, and Building (Kindle) · davidbellairian.com
Comments