Ask an AI to analyze a rental deal and it will produce something impressive in thirty seconds: cap rate, rent assumptions, a rehab budget, an exit price. It reads like underwriting. Some of it is arithmetic on numbers you gave it, which is trustworthy. The rest is the model filling gaps with plausible-sounding inventions, and it will not tell you which parts are which. Knowing the difference is the whole skill of using AI on deals.
The short version
AI is strong at reading what you hand it and weak at knowing what you did not. Feed it a lease stack, a rent roll, and a settlement statement and it will extract, summarize, and cross-check faster than you can. Ask it for an ARV, a rehab number on a house it has never seen, or current rents in a specific submarket, and you are collecting confident fiction. The models with the best data in history lost real money on exactly this problem.
What AI does genuinely well on a deal
The pattern behind every good use: the facts come from you, the model does the reading.
- Lease extraction. Hand it a stack of leases and get terms, escalations, deposits, renewal options, and unusual clauses in a table. Then spot-check a sample against the documents, because extraction is strong but not perfect.
- Rent roll sanity checks. Models are good at internal consistency: unit counts that do not sum, escalation dates that contradict lease terms, deposits that do not match. Finding the seams inside documents is real work it does cheaply.
- Settlement statement and payoff review. Summarizing where every dollar went, and flagging line items that look off against the contract.
- Scenario grinding. Once your assumptions are pinned down, running them across price points and rate cases is arithmetic, and tools that compute rather than freestyle are reliable at it.
- First-pass document translation. Survey exceptions, title commitments, HOA resale packets: a plain-language summary tells you where to spend your careful reading. It does not replace the careful reading.
Where AI will quietly lie to you
Language models fabricate specifics with full confidence. The famous courtroom version of this involved lawyers sanctioned for filing briefs citing cases the AI had invented outright. The real estate version is subtler because the fabrications are plausible: a rent comp that sounds right for the neighborhood, a price per square foot that fits the vibe, a repair estimate with satisfying line items. None of it came from anywhere.
Three specific traps for investors:
ARV and comps. A model without live local data cannot know what closed last month, and in Texas it is worse: sale prices are not public record here, so even data-connected tools work from thinner inputs than in most states. An ARV from a chat window is a guess wearing a spreadsheet costume.
Rehab numbers sight unseen. Condition, foundation, sewer, roof age, and what is behind the drywall exist in no dataset. Even the best-funded valuation models, with better data than any retail tool, produced errors large enough that their owner shut down a home-buying business built on them. Your contractor walking the property is not a step AI compresses.
“Current” anything. Models have knowledge cutoffs. Rates, rents, insurance quotes, and inventory move faster than model training. Anything time-sensitive must come from a live source, and the model should be told to use what you gave it rather than its memory.
The discipline that fixes all three is the same: before you act on an AI-assisted analysis, list every number in it and mark where each one came from. Anything the model supplied on its own gets replaced with a sourced figure or thrown out.
A note on rent-setting tools
Underwriting a purchase with your own comps is one thing. Ongoing pricing tools that pool competitors’ private data are another: that model drew a federal antitrust action against its biggest vendor, and a growing list of states and cities ban algorithmic rent-setting tools outright. Texas and Oklahoma currently do not, but a small landlord pricing off public comps loses nothing by staying on the safe side of that line.
Frequently asked questions
Can AI estimate repair costs from listing photos?
It can produce an estimate, which is different from a useful one. Photos are marketing, increasingly edited marketing, and the expensive problems are the ones photos do not show. Use AI to build your scope-of-work checklist; use a walkthrough and a contractor to price it.
Is an online value estimate good enough for an offer?
For screening which deals deserve attention, yes. For the offer, no, and doubly so in Texas where the models cannot see closed prices. The published accuracy figures for homes not currently listed are far looser than most investors assume, and half of estimates miss by more than the headline number.
What is the single best AI habit for underwriting?
Make the model show sources. Ask it to separate what came from your documents from what it assumed. A model that has to label its assumptions is a model you can audit, and the labeling request costs nothing.
Where we land on it
We use AI on deals the way we would use a sharp intern: read everything, summarize, flag inconsistencies, and never sign anything on its numbers alone. The dollar-shaped decisions still run on walked properties, sourced comps, and contractor bids. If you are investing in DFW or the OKC corridor and want deal flow underwritten by people who stand in the houses, start here, and our post on AI tenant screening covers the operating side of the same discipline.
Enjoying these? We publish straight talk like this twice a week. Follow us on Linktree to keep up with everything we are working on, from new tools to new markets.